Asesoramiento electrónico sobre la dosis de los fármacos para mejorar la práctica de la prescripción
Resumen
Antecedentes
El mantenimiento de las concentraciones terapéuticas de los fármacos con una ventana terapéutica estrecha es una tarea compleja. Se han diseñado varios sistemas de computación para ayudar a los médicos a determinar la dosis óptima del fármaco. Se podrían lograr mejorías significativas en la asistencia sanitaria si el asesoramiento electrónico mejorara los resultados de salud y si se pudiera ejecutar en la práctica habitual de una manera rentable. Ésta es una versión actualizada de una revisión sistemática Cochrane anterior, publicada por primera vez en 2001 y actualizada en 2008.
Objetivos
Evaluar si el asesoramiento electrónico sobre la dosis del fármaco presenta efectos beneficiosos sobre los resultados de los pacientes en comparación con la atención habitual (dosificación empírica sin asistencia electrónica).
Métodos de búsqueda
Se hicieron búsquedas en las siguientes bases de datos desde 1996 hasta enero 2012: registro especializado del Grupo EPOC, Reference Manager; Registro Cochrane Central de Ensayos Controlados (Cochrane Central Register of Controlled Trials) (CENTRAL), MEDLINE, Ovid; EMBASE (Ovid; y CINAHL, EbscoHost. Se realizó una búsqueda de actualización para el período enero 2012 hasta enero 2013; estos resultados fueron examinados por los autores y los estudios potencialmente relevantes se enumeran en "Estudios en espera de clasificación". Los autores de la revisión también buscaron las listas de referencias de los estudios pertinentes y las revisiones relacionadas.
Criterios de selección
Se incluyeron ensayos controlados aleatorios, estudios controlados no aleatorios, estudios controlados del tipo antes y después (before and after studies) y análisis de series de tiempo interrumpido del asesoramiento electrónico sobre la dosis de fármacos. Los participantes fueron profesionales sanitarios responsables del cuidado de los pacientes. Los resultados fueron cualquier cambio en la salud de los pacientes medido de forma objetiva y que fuese resultado del asesoramiento electrónico (como el control de los fármacos del tratamiento, la mejoría clínica, las reacciones adversas).
Obtención y análisis de los datos
Dos autores de la revisión, de forma independiente, extrajeron los datos y evaluaron la calidad de los estudios. Los resultados de los estudios incluidos se agruparon por fármaco utilizado y por efecto deseado para los antibióticos aminoglucósidos, la amitriptilina, los anestésicos, la insulina, los anticoagulantes, la estimulación ovárica, los fármacos antirrechazo y la teofilina. Los tamaños del efecto se combinaron para proporcionar un efecto general para cada subgrupo de estudios, mediante un modelo de efectos aleatorios. Cuando fue apropiado, los estudios se agruparon de forma adicional por tipo de resultado (es decir, cuando no hubo pruebas de heterogeneidad).
Resultados principales
Se incluyeron 46 comparaciones (de 42 ensayos) (en comparación con 26 comparaciones de la última actualización) con una gama amplia de fármacos en ámbitos hospitalarios y ambulatorios. Todos fueron ensayos controlados aleatorios, excepto dos estudios. Las intervenciones generalmente se dirigieron a los médicos, aunque algunos estudios intentaron influir en las prescripciones de farmacéuticos y enfermeras. Los fármacos evaluados fueron anticoagulantes, insulina, antibióticos aminoglucósidos, teofilina, fármacos antirrechazo, agentes anestésicos, antidepresivos y gonadotropinas. Aunque todos los estudios utilizaron medidas de resultado confiables, su calidad fue generalmente deficiente.
Esta actualización encontró resultados similares a la actualización anterior y logró identificar áreas terapéuticas específicas en las que el asesoramiento electrónico sobre la dosis del fármaco fue beneficioso en comparación con la atención habitual:
1. aumentó las concentraciones séricas máximas proyectadas (diferencia de medias estandarizada [DME] 0,79; IC del 95%: 0,46 a 1,13) y la proporción de pacientes con concentraciones del fármaco en plasma dentro del rango terapéutico después de dos días (cociente de riesgos [CR] agrupado 4,44; IC del 95%: 1,94 a 10,13) para los antibióticos aminoglucósidos;
2. dio lugar a un parámetro fisiológico que frecuentemente estuvo dentro del rango deseado para los anticoagulantes orales (DME para el porcentaje de tiempo en la razón internacional normalizada proyectada +0,19; IC del 95%: 0,06 a 0,33) y la insulina (DME para el porcentaje de tiempo en el rango de la glucosa proyectada: +1,27; IC del 95%: 0,56 a 1,98);
3. disminuyó el tiempo hasta lograr la estabilización para los anticoagulantes orales (DME ‐0,56; IC del 95%: ‐1,07 a ‐0,04);
4. redujo los eventos de tromboembolia (cociente de tasas 0,68; IC del 95%: 0,49 a 0,94) y tendió a reducir los eventos hemorrágicos para los anticoagulantes, aunque la diferencia no fue significativa (cociente de tasas 0,81; IC del 95%: 0,60 a 1,08). Tendió a reducir los efectos no deseados para los antibióticos aminoglucósidos (nefrotoxicidad: CR 0,67; IC del 95%: 0,42 a 1,06) y los fármacos antirrechazo (infecciones por citomegalovirus: CR 0,90; IC del 95%: 0,58 a 1,40);
5. tendió a reducir la duración de la estancia en el hospital, aunque la diferencia no fue significativa (DME ‐0,15; IC del 95%: ‐0,33 a 0,02) y a lograr razones de costo‐eficacia equivalentes o mejores que la atención habitual;
6. no hubo pruebas de diferencias en la mortalidad u otros eventos adversos clínicos para la insulina (hipoglucemia), los agentes anestésicos, los fármacos antirrechazo y los antidepresivos.
Para todos los resultados, la heterogeneidad estadística cuantificada por las estadísticas de I2 fue moderada a alta.
Conclusiones de los autores
Esta actualización de la revisión indica que el asesoramiento electrónico para la dosis del fármaco tiene algunos beneficios: aumenta las concentraciones séricas para los antibióticos aminoglucósidos y mejora la proporción de pacientes para los cuales el nivel del fármaco en plasma está dentro del rango terapéutico para los antibióticos aminoglucósidos.
Da lugar a un parámetro fisiológico que frecuentemente está dentro del rango deseado para los anticoagulantes orales y la insulina. Disminuye el tiempo hasta lograr la estabilización para los anticoagulantes orales. Tiende a reducir los efectos no deseados para los antibióticos aminoglucósidos y los fármacos antirrechazo, y reduce significativamente los eventos de tromboembolia para los anticoagulantes. Tiende a reducir la duración de la estancia hospitalaria en comparación con la atención habitual, aunque se lograron cocientes de costo‐eficacia equivalentes o mejores.
Sin embargo, no hubo pruebas de que el apoyo a las decisiones tuviese un efecto sobre la mortalidad u otros eventos adversos clínicos para la insulina (hipoglucemia), los agentes anestésicos, los fármacos antirrechazo y los antidepresivos. Además, no hubo pruebas que indiquen que algunas características técnicas que apoyan la toma de decisiones (como la integración en un sistema electrónico de ingreso de las indicaciones médicas) o los aspectos de la organización de la atención (como el ámbito) pudieran optimizar el efecto del asesoramiento electrónico.
Teniendo en cuenta el alto riesgo de sesgo y alta heterogeneidad entre los estudios, estos resultados deben interpretarse con precaución.
PICO
Resumen en términos sencillos
Asesoramiento electrónico sobre la dosis de los fármacos para mejorar la práctica de la prescripción
Antecedentes
A menudo, los médicos y otros profesionales sanitarios prescriben fármacos que sólo producirán efecto en ciertas concentraciones. Se considera que estos fármacos tienen una ventana terapéutica estrecha. Lo anterior indica que si el nivel del fármaco es demasiado alto o demasiado bajo, puede provocar efectos secundarios graves o no proporcionar los beneficios que debería. Por ejemplo, los fármacos que disminuyen la viscosidad de la sangre (anticoagulantes) se prescriben para disminuir su viscosidad y prevenir los coágulos. Si la concentración es demasiado alta, los pacientes pueden experimentar hemorragia excesiva e incluso la muerte. Por el contrario, si la concentración es demasiado baja, podría formarse un coágulo que causaría un accidente cerebrovascular. Para estos tipos de fármacos, es importante que se prescriba la cantidad correcta del fármaco.
El cálculo y la prescripción de la cantidad correcta pueden ser complicados y requerir mucho tiempo para los profesionales de la atención sanitaria. A veces, la determinación de la dosis correcta puede tomar mucho tiempo debido a que los profesionales de la asistencia sanitaria pueden no desear prescribir dosis altas de los fármacos al inicio porque cometen errores en los cálculos. Se han diseñado varios sistemas electrónicos para hacer estos cálculos y ayudar a los profesionales de la salud a prescribir este tipo de fármacos.
Características de los estudios
Se buscaron pruebas de ensayos clínicos en las bases de datos científicas para evaluar la efectividad de estos sistemas electrónicos. Las pruebas se actualizaron hasta enero de 2012. Se encontraron datos de 42 ensayos (40 ensayos controlados aleatorios [ensayos que asignan a los pacientes al azar para recibir uno entre varios fármacos o procedimientos] y dos ensayos controlados no aleatorios).
Resultados clave
El asesoramiento electrónico en cuanto a la dosis del fármaco puede beneficiar a los pacientes que reciben determinados fármacos en comparación con la dosificación empírica (cuando la dosis se elige sobre la base de las observaciones y la experiencia del médico) sin asesoramiento electrónico. Al utilizar el sistema electrónico, los profesionales de la asistencia sanitaria prescribieron dosis mayores apropiadas de los fármacos al inicio para los antibióticos aminoglucósidos y la dosis correcta del fármaco se alcanzó más rápidamente para los anticoagulantes orales. La utilización de dicho sistema redujo significativamente los eventos de tromboembolia (coágulos de sangre) para los anticoagulantes y tendió a reducir los efectos no deseados para los antibióticos aminoglucósidos y los fármacos antirrechazo (aunque no se observó una diferencia importante). Tendió a reducir la duración de la estancia hospitalaria en comparación con la atención habitual con un costo‐eficacia equivalente o mejor. No hubo pruebas de efectos sobre la muerte o los eventos secundarios clínicos para la insulina (nivel bajo de azúcar en sangre [hipoglucemia]), los agentes anestésicos, los fármacos antirrechazo (fármacos administrados para prevenir el rechazo de un órgano trasplantado) y los antidepresivos.
Calidad de la evidencia
La calidad de los estudios fue baja por lo que estos resultados deben interpretarse con precaución.
Authors' conclusions
Summary of findings
| Computerized advice on drug dosage for leading serum concentrations within therapeutic range | ||||||
| Patient or population: patients with leading serum concentrations within therapeutic range | ||||||
| Outcomes | Illustrative comparative risks* (95% CI) | Relative effect | No of participants | Quality of the evidence | Comments | |
| Assumed risk | Corresponding risk | |||||
| Control | Computerized advice on drug dosage | |||||
| Serum concentrations ‐ aminoglycoside antibiotics: peak concentration | ‐ | The mean serum concentrations ‐ aminoglycoside antibiotics: peak concentration in the intervention groups was | ‐ | 372 | ⊕⊕⊝⊝ | SMD 0.79 (95% CI 0.46 to 1.13) |
| Serum concentrations ‐ theophylline | ‐ | The mean serum concentrations ‐ theophylline in the intervention groups was | ‐ | 201 | ⊕⊕⊝⊝ | SMD 0.41 (95% CI ‐0.2 to 1.02) |
| Proportion of participants within therapeutic range ‐ aminoglycoside antibiotics: % of participants with peak concentrations adequate after 2 days | Study population | RR 4.44 | 72 | ⊕⊕⊕⊝ | ‐ | |
| 135 per 1000 | 600 per 1000 | |||||
| Moderate | ||||||
| 151 per 1000 | 670 per 1000 | |||||
| Proportion of participants with toxic drug levels ‐ theophylline | Study population | RR 0.53 | 109 | ⊕⊕⊕⊝ | ‐ | |
| 273 per 1000 | 145 per 1000 | |||||
| Moderate | ||||||
| 217 per 1000 | 115 per 1000 | |||||
| *The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). | ||||||
| GRADE Working Group grades of evidence | ||||||
| 1 Lack of blinding of participants and personnel in all studies. Incomplete outcome data in three studies. Random sequence generation and allocation concealment unclear in one study. | ||||||
| Computerized advice on drug dosage for leading physiological parameters within therapeutic range | ||||||
| Patient or population: patients with leading physiological parameters within therapeutic range | ||||||
| Outcomes | Illustrative comparative risks* (95% CI) | Relative effect | No of participants | Quality of the evidence | Comments | |
| Assumed risk | Corresponding risk | |||||
| Control | Computerized advice on drug dosage | |||||
| Oral anticoagulants: time in target INR range (%) | ‐ | The mean oral anticoagulants: time in target INR range (%) in the intervention groups was | ‐ | 13,581 | ⊕⊝⊝⊝ | SMD 0.19 (95% CI 0.06 to 0.33) |
| Insulin: time in target glucose range (%) | ‐ | The mean insulin: time in target glucose range (%) in the intervention groups was | ‐ | 234 | ⊕⊕⊝⊝ | SMD 1.27 (95% CI 0.56 to 1.98) |
| *The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). | ||||||
| GRADE Working Group grades of evidence | ||||||
| 1 No information given on random sequence generation and allocation concealment in half of the studies. | ||||||
| Computerized advice on drug dosage for reducing time to achieve therapeutic control | ||||||
| Patient or population: patients with reducing time to achieve therapeutic control | ||||||
| Outcomes | Illustrative comparative risks* (95% CI) | Relative effect | No of participants | Quality of the evidence | Comments | |
| Assumed risk | Corresponding risk | |||||
| Control | Computerized advice on drug dosage | |||||
| Time to achieve therapeutic range ‐ oral anticoagulants: time to achieve therapeutic prothrombin ratio (days) | ‐ | The mean time to achieve therapeutic range ‐ oral anticoagulants: time to achieve therapeutic prothrombin ratio (days) in the intervention groups was | ‐ | 223 | ⊕⊕⊝⊝ | SMD ‐0.22 (95% CI ‐0.69 to 0.26) |
| Time to achieve therapeutic range ‐ insulin: time to achieve therapeutic control (hours) | ‐ | The mean time to achieve therapeutic range ‐ insulin: time to achieve therapeutic control (hours) in the intervention groups was | ‐ | 194 | ⊕⊝⊝⊝ | SMD ‐0.14 (95% CI ‐0.98 to 0.7) |
| Time to stabilization ‐ oral anticoagulants: time to stabilization (days) | ‐ | The mean time to stabilization ‐ oral anticoagulants: time to stabilization (days) in the intervention groups was | ‐ | 255 | ⊕⊝⊝⊝ | SMD ‐0.56 (95% CI ‐1.07 to ‐0.04) |
| *The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). | ||||||
| GRADE Working Group grades of evidence | ||||||
| 1 Allocation concealment was unclear, and there was no blinding of participants and personnel in both studies. | ||||||
| Computerized advice on drug dosage for leading to fewer clinical adverse events | ||||||
| Patient or population: patients with leading to fewer clinical adverse events | ||||||
| Outcomes | Illustrative comparative risks* (95% CI) | Relative effect | No of participants | Quality of the evidence | Comments | |
| Assumed risk | Corresponding risk | |||||
| Control | Computerized advice on drug dosage | |||||
| Death | Study population | RR 1.08 | 14,046 | ⊕⊕⊝⊝ | ‐ | |
| 12 per 1000 | 13 per 1000 | |||||
| Moderate | ||||||
| 28 per 1000 | 30 per 1000 | |||||
| Anticoagulants: events ‐ bleeding | Study population | RR 0.65 | 552 | ⊕⊕⊝⊝ | ‐ | |
| 61 per 1000 | 40 per 1000 | |||||
| Moderate | ||||||
| 65 per 1000 | 42 per 1000 | |||||
| Anticoagulants: events ‐ thromboembolism | Study population | RR 3.25 | 355 | ⊕⊝⊝⊝ | ‐ | |
| 11 per 1000 | 36 per 1000 | |||||
| Moderate | ||||||
| 7 per 1000 | 23 per 1000 | |||||
| Insulin ‐ hypoglycaemia (< 60 mg/dL) | Study population | RR 0.71 | 378 | ⊕⊕⊝⊝ | ‐ | |
| 90 per 1000 | 64 per 1000 | |||||
| Moderate | ||||||
| 67 per 1000 | 48 per 1000 | |||||
| Insulin ‐ severe hypoglycaemia (< 40 mg/dL) | Study population | RR 0.69 | 292 | ⊕⊝⊝⊝ | ‐ | |
| 14 per 1000 | 9 per 1000 | |||||
| Moderate | ||||||
| 8 per 1000 | 6 per 1000 | |||||
| Aminoglycoside antibiotics ‐ nephrotoxicity | Study population | RR 0.67 | 493 | ⊕⊕⊝⊝ | ‐ | |
| 162 per 1000 | 108 per 1000 | |||||
| Moderate | ||||||
| 154 per 1000 | 103 per 1000 | |||||
| *The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). | ||||||
| GRADE Working Group grades of evidence | ||||||
| 1 No (or unclear) blinding of participants and personnel in all studies. In half of the studies: sequence generation and/or allocation concealment were unclear, and the data were incomplete. | ||||||
| Saving healthcare resources for saving healthcare resources | ||||||
| Patient or population: saving healthcare resources | ||||||
| Outcomes | Illustrative comparative risks* (95% CI) | Relative effect | No of participants | Quality of the evidence | Comments | |
| Assumed risk | Corresponding risk | |||||
| Control | Saving healthcare resources | |||||
| Length of stay (days) | ‐ | The mean length of stay (days) in the intervention groups was | ‐ | 18,507 | ⊕⊝⊝⊝ | SMD ‐0.15 (95% CI ‐0.33 to 0.02) |
| Length of stay (days) ‐ oral anticoagulants | ‐ | The mean length of stay (days) ‐ oral anticoagulants in the intervention groups was | ‐ | 105 | ⊕⊝⊝⊝ | SMD ‐0.12 (95% CI ‐1.1 to 0.86) |
| Length of stay (days) ‐ insulin | ‐ | The mean length of stay (days) ‐ insulin in the intervention groups was | ‐ | 128 | ⊕⊕⊕⊕ | SMD 0.18 (95% CI ‐0.17 to 0.53) |
| Length of stay (days) ‐ theophylline | ‐ | The mean length of stay (days) ‐ theophylline in the intervention groups was | ‐ | 151 | ⊕⊕⊝⊝ | SMD ‐0.2 (95% CI ‐0.56 to 0.16) |
| Length of stay (days) ‐ aminoglycoside antibiotics | ‐ | The mean length of stay (days) ‐ aminoglycoside antibiotics in the intervention groups was | ‐ | 295 | ⊕⊕⊕⊝ | SMD ‐0.35 (95% CI ‐0.58 to ‐0.12) |
| Length of stay (days) ‐ anti‐rejection drugs | ‐ | The mean length of stay (days) ‐ anti‐rejection drugs in the intervention groups was | ‐ | 17,828 | ⊕⊕⊕⊝ | SMD ‐0.04 (95% CI ‐0.07 to ‐0.01) |
| *The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). | ||||||
| GRADE Working Group grades of evidence | ||||||
| 1 I2 = 57%. | ||||||
Background
Medication mistakes still represent 20% of all medical errors although many efforts have focused in recent years on reducing the risk of drug errors (Brennan 1991). Maintaining therapeutic concentrations of drugs is a complex task requiring knowledge of evidence‐based clinical guidelines, clinical pharmacology and skills in dose calculation. The potential for error is great since many of the drugs commonly used have a narrow 'window' within which therapeutic benefits can be obtained with a low risk of unwanted effects.
Description of the condition
Monitoring drug therapy to optimize effects and minimize dangers can be very time consuming. Practitioners may need access to a large amount of information to make an appropriate dose adjustment in situations such as prevention of deep vein thrombosis or management of people with renal insufficiency (Durieux 2005). Under these conditions, healthcare professionals make errors of judgement because their ability to process information is finite (McDonald 1976a).
For example, in ambulatory settings, general practitioners (GP) reported difficulties with drug dosing, especially for children, elderly people and people with renal impairment (Franke 2000). Moreover, physicians' computational abilities are often insufficient to perform calculations for drug dosage (Baldwin 1995). For example, 82 out of 150 hospital doctors were unable to calculate how many milligrams of lidocaine were in a 10 mL ampoule of 1% solution (Rolfe 1995).
Description of the intervention
Clinical decision support systems (CDSS), either computerized or not, have been proposed to improve clinical practice (Kawamoto 2005). Ideally, decision support, integrated in the electronic medical record as the platform, can provide physicians with tools making it possible to improve practice and patient safety (Bates 2003). An effective decision support system would anticipate needs and deliver information quickly in real time, adapted to the user's workflow (Bates 2003).
How the intervention might work
Computers are very good at collecting information and performing repetitive calculations. Moreover, the drugs that cause the most problems have often been in use for many years. The pharmacology of the drugs is, therefore, well understood and thus computer models can be used to generate advice on dosage. Several types of computer systems have been designed to help doctors in the task of determining the optimum dosage of drugs. Significant improvements in health could be achieved if computer advice was shown to be beneficial and was provided by the computers that clinicians now use for their everyday work.
In addition, the logistics by which the advice on drug dosage is delivered to the healthcare professional is critical to its effectiveness and to the transferability of this effectiveness in other settings. Computer physician order entry (CPOE) systems, which allow physicians to enter orders directly into a computer rather than hand writing them, have the potential to incorporate CDSS into daily practice (Kuperman 2003). According to three systematic reviews (Garg 2005; Kawamoto 2005; Nieuwlaat 2011a), CDSSs are more often associated with improvement of practice when the decision aid is automatically prompted, integrated in clinicians' workflow, and provided at time and location of decision making.
This review focuses on advice with a personalized dosage for a specific participant. Two other Cochrane systematic reviews were interested in computer‐generated reminders: one evaluated the effects of on‐screen computer reminders delivered to clinicians at the point of care (Shojania 2009), and one considered reminders delivered on paper to healthcare professionals (Arditi 2012).
Why it is important to do this review
This is an updated version of two earlier Cochrane systematic reviews (Walton 2001; Durieux 2008).Those earlier reviews provided some evidence to support the use of computer assistance in determining drug dosage but concluded that further clinical trials were necessary to confirm those results.
Objectives
The primary objective of this systematic review was to examine whether computerized advice on drug dosage given to healthcare professionals is beneficial when compared with routine care (empiric dosing without computer assistance). The secondary objective was to determine whether any technical features of computerized systems or organizational aspects concerning their implementation influence their effectiveness.
Hypothesis tested
In previous versions of this review, we tested the benefit in terms of effects on process of care (healthcare professional oriented) and on outcome of care (patient oriented). The effect on process of care was any change in drug dose as a process measure. However, there are problems with this approach since a higher dose may, in some circumstances, be beneficial and, in others, be disadvantageous. Thus, in this review we have removed this outcome and only reported effects on dosage where we judged the changes to be meaningful. We examined six hypotheses on patient‐oriented outcomes (most of them measuring surrogate outcomes).
Effect on outcome of care (patient‐oriented outcomes)
-
Decisions on drug dosage based on computer advice lead more often to drug levels within the therapeutic range.
-
Decisions on drug dosage based on computer advice lead more often to a physiological parameter being maintained within the desired range (e.g. blood pressure or prothrombin (PT) time).
-
Decisions on drug dosage based on computer advice lead to more rapid therapeutic control, assessed by a physiological parameter.
-
Decisions on drug dosage based on computer advice lead to greater clinical effectiveness, assessed by clinical improvement.
-
Decisions on drug dosage based on computer advice lead to fewer unwanted effects than conventional dose adjustment.
-
Computer advice reduces the cost of health care or the use of resources (e.g. length of hospital stay).
Unlike the previous update of this review, we decided to analyze the results individually for each drug because of a high clinical and statistical heterogeneity between drugs (each drug has its specific outcomes).
Other hypotheses address our secondary objectives and reflect a series of subgroup analyses.
Effect of decision support logistics and organization of care
-
Computer advice given in real time is more effective than that given by delayed feedback.
-
Computer advice integrated in CPOE system is more effective than other systems.
-
System‐initiated computer advice is more effective than user‐initiated computer advice.
-
Direct intervention (system delivers advice directly to the provider) is more effective than indirect intervention (advice is made available to the provider by the intermediate of a third party actor, i.e. system is not directly used by the provider).
-
The impact of computer advice depends on the setting where it is implemented (inpatient versus outpatient care).
-
Computer advice given as a recommendation is more effective than a calculated dose proposed without possibility of change and which does not take into account the healthcare professional's experience).
Methods
Criteria for considering studies for this review
Types of studies
We considered for inclusion all types of study designs that met Effective Practice and Organization of Care Group (EPOC) inclusion criteria:
-
Randomized controlled trials (RCTs) where the unit of randomization was:
-
the participant or
-
the cluster: healthcare professionals (doctors, nurses or pharmacists) or groups of professionals (practices or hospitals);
-
Non‐randomized controlled trials (NRCTs);
-
Controlled before‐and‐after (CBA) studies with:
-
the same pre‐ and postintervention periods for study and control sites;
-
comparable study and control sites with respect to level of care, setting of care and baseline characteristics;
-
two intervention sites and two control sites;
-
Interrupted time series (ITS) studies with:
-
a clearly defined point in time when the intervention occurred;
-
at least three data points before and three after the intervention.
See the EPOC checklist for definition of designs (EPOC 2012).
Types of participants
The participants were healthcare professionals with responsibility for patient care.
Types of interventions
We sought to identify all comparative studies comparing computerized advice on drug dosage given to routine care (empiric dosing without computer assistance). We defined computerized advice on drug dosage as a recommendation provided to the healthcare professional on the drug dosage needed for a specific participant and a specific drug and calculated by a computer.
Computer program
The computer program was a software model or an application integrated into a laptop, a smartphone, a tablet computer, the CPOE, or a website (online calculator). We did not consider interventions where the recommended drug dose was not calculated by a computer, for example an equation or a nomogram not implemented in a computer device.
Nature of advice
The advice included a dosage personalized for a specific participant. We did not include studies reporting a popup with general advice on the dosage required for a specific condition (most frequently dose of medication, dose interval, maximum total daily dose).
Content of recommendation and calculation of drug dose
The recommendation could be evidence‐based, a clinical practice guideline developed by expert bodies (government, professional) or local clinicians, or population pharmacokinetic/pharmacodynamic models. The participant's drug dose was computed using an equation including participant's characteristics (participant's age, weight, previous drug levels…). Alternatively, a more complex mathematical model was used, which was generally a pharmacokinetic model of the relationship between administered doses of the drug and observed concentrations in the participant's body.
Advice delivery and timing
The computerized advice could be delivered to clinicians when they are writing their prescription (point of care delivery) or it could be delivered at a later time. In addition, the computerized advice could be delivered to another healthcare professional (e.g. a pharmacist or a pharmacokinetic unit) and passed to the clinician. Unlike the review on the effect of point‐of‐care computer reminders on physician behaviour (Shojania 2009), we included systems that were not encountered during routine performance of the activities of interest, for example a dedicated computer used only for performing dose calculation for anticoagulants. These systems require clinicians to depart from their usual workflow in order to avail themselves of the reminder or decision support.
Control of the healthcare professional
We included studies where advice was given as a recommendation so that the healthcare professional was able to accept or refuse it. We did not include studies reporting non‐specific advice given to a healthcare professional to adjust drug dosage or when the healthcare professional was not in charge of every adjustment of the drug, for example studies reporting the direct administration of a drug to the participant by means of a computer supervised device (closed‐loop system) or through self dosing devices (one Cochrane review addresses the evaluation of anticoagulant self management (Garcia‐Alamino 2010)).
Starter
The advice could be system‐initiated (advice appears without user intervention) or user‐initiated (the computer program must be started by the user to obtain an advice).
Control group
The control group included empiric dosing without computer assistance, in general routine care.
Types of outcome measures
All outcome measures included were patient‐oriented outcomes.
-
Proportion of participants or time for which the plasma drug concentrations was within the therapeutic range.
-
Proportion of participants or time for which the studied physiological parameter was maintained within the target range.
-
Time to achieve therapeutic control.
-
Proportion of participants with toxic drug levels.
-
Proportion of participants with clinical improvement.
-
Proportion of participants with adverse effects of drug therapy.
-
Proportion of deaths.
-
Length of hospital stay.
-
Total cost per participant.
We excluded the outcomes for which reporting was incomplete (e.g. no numerical values reported, no measure of dispersion) and excluded studies whose only relevant outcomes were death or adverse effects requiring monitoring and which did not explicitly report such outcomes as primary.
For serum concentrations, we also considered peak, trough and steady‐state concentrations. The measurement of drug levels in the blood, called therapeutic drug monitoring (TDM), is required for some drugs to ensure that the participants maintain the concentration of drug within the established therapeutic range (drug effective without toxicity). Blood for peak level is collected at the drug's highest concentration within the dosing period. Trough levels (occasionally called residual levels) are measured just prior to administration of the next dose, and are the lowest concentration in the dosing interval. Most therapeutic drugs have a narrow trough to peak difference (therapeutic range), and, therefore, only trough levels are needed to detect blood levels that are too low or too high. Peak levels are needed for some drugs, especially aminoglycoside antibiotics: a concentration below the therapeutic range will not resolve the bacterial infection so high peak concentrations are necessary for optimal efficacy; however, too high a level can cause damage such as nephrotoxicity so it is important that the trough concentration be allowed to fall in order to avoid accumulation. The steady‐state concentration is defined as the point at which the amount of drug administered (drug intake) and the amount of drug excreted (drug elimination) reach an equilibrium. The goal of TDM is to optimize the drug dose so that the participant's drug concentrations remain within the therapeutic range.
Search methods for identification of studies
Updates to Cochrane systematic reviews usually entail executing previously used search strategies for the update period or, in other words, from the date of the last search, to present. In some cases, however, search strategies must be assessed and revised in order to optimize the identification of evidence. When this occurs, it is advisable to search retrospectively‐‐that is, to re‐search previously searched time periods, in order to discover whether or not studies have been missed. This update represents a review where search strategies have been revised significantly (by M. Fiander, Information Scientist, and Trials Search Co‐ordinator for the EPOC Group) and where, consequently, searching has been conducted not only from the date of last search in 2006, but retrospectively from 1996 to January 2012, where 1996 to 2006 represents a previously searched time period.
The revised search strategies had an impact on the review since the strategies identified a number of studies which should have been found during previous searches: Ageno 2000, Claes 2005, Claes 2006, Mitra 2005, Plank 2006, Poller 2002, Poller 2003.
For the initial review, the databases listed below were searched from database start date to 1996; for the first update, the search period was 1996 to 2006; for this, the second, update, searches were run from 1996 to January 2012. Two methodological search filters were used to limit retrieval to appropriate study designs: the Cochrane Highly Sensitive Search Strategy (sensitivity‐ and precision‐maximizing version, 2008 revision) to identify randomized trials; and an EPOC methodology filter (Appendix 1) to identify non‐RCT designs. Related reviews were identified by searching the Database of Abstracts of Reviews (DARE). All databases were searched from their start date forward; start date represents the date the database began to index journals. Note that database start dates vary by database (Medline, EMBASE, etc.) and provider (OVID, Ebsco, etc.). Start dates are provided in the list of databases, below.
A top‐up search was conducted for the period January 2012 to January 2013; the authors screened the titles and abstracts of these results and added potentially relevant studies to Studies awaiting classification.
The original MEDLINE search strategies used until 2006 are in Appendix 2. The revised search strategies used for this update are in Appendices 2 to 6 as follows: MEDLINE, Appendix 3; EMBASE, Appendix 4; CINAHL, Appendix 5; Cochrane Central Register of Controlled Trials, Appendix 6; EPOC Specialised Register, Appendix 7.
Databases searched
-
Cochrane Central Register of Controlled Trials (CENTRAL), Issue 12, 2012, OvidSP EBM Reviews
-
MEDLINE, including In‐Process & Other Non‐Indexed Citations, OvidSP, 1946‐January 2013
-
EMBASE, 1947 to January 2013, OvidSP
-
CINAHL (Cumulative Index to Nursing and Allied Health Literature), 1980‐January 2013, EbscoHost
-
EPOC Group, Specialised Register
Searching other resources
The review authors handsearched reference lists from primary articles and relevant reviews identified, and conference proceedings. We contacted experts in the field.
We also:
-
Screened (hand searched) the following journals:
-
Therapeutic Drug Monitoring journal (1979 to December 2006).
-
Journal of the American Medical Informatics Association (January 1996 to March 2007).
-
-
Reviewed reference lists of all included studies, relevant systematic reviews and primary studies.
-
Contacted authors of relevant studies/ reviews to clarify reported published information and to seek unpublished results/data.
-
Contacted researchers with expertise relevant to the review topic/ EPOC interventions.
Data collection and analysis
Selection of studies
We merged the search results using Reference Manager 5 (RevMan 2011), and removed duplicate records. We examined all titles and abstracts.
Two review authors (FG, PD) examined independently each title and abstract to exclude obviously irrelevant reports (mainly therapeutic trials and genetic research). We retrieved full texts, which were independently screened. We then randomly allocated each selected study to two pairs of review authors (IC and FG, MR and PD) who reviewed it and extracted data independently. We resolved disagreements by group discussion with the four review authors. We reported reasons for excluding full papers.
Data extraction and management
We reviewed the data abstraction form for the previous update of the review. We adapted a checklist to the specific subject to extract the decision support technical features by which the advice on drug dosage was delivered to the healthcare professional.
-
Was the computerized advice delivered in real time (at the moment of the practitioners decision making) or by delayed feedback?
-
Was the computerized advice integrated in a CPOE?
-
Was the computerized advice user‐initiated or system‐initiated?
-
Was the intervention direct or indirect (a third party brought advice from computer and transfers it to user)?
An additional feature was added:
-
Was the information of the calculated dose given as a recommendation to the healthcare professional who prescribed through the computer or through another healthcare professional (the healthcare professional had the possibility to accept or refuse the advice)?
The review authors abstracted the data independently and resolved disagreements by discussion. A statistician (FG) reviewed all data and contacted authors of included studies for additional information.
Assessment of risk of bias in included studies
We assessed the risk of bias of the studies using the 'Risk of bias' criteria described by the EPOC group and extracted data using the EPOC checklist (EPOC 2009; Higgins 2011). We used nine standard criteria for RCTs, NRCTs and CBAs: allocation sequence adequately generated; allocation adequately concealed; baseline outcome measurements similar; baseline characteristics similar; incomplete outcome data adequately addressed; knowledge of the allocated interventions adequately prevented during the study; study adequately protected against contamination; study free from selective outcome reporting; study free from other risks of bias. We scored studies using cluster randomization to be adequate on concealment of allocation (if the sequence generation was adequate) and on protection against contamination. Baseline characteristics were considered for similarity at the unit of analysis level. Risk of bias on baseline outcome measurements was only evaluated for insulin. For the other drugs, the baseline measurement was not relevant since there was no drug intake before the intervention. Therefore, we considered risk of bias on baseline outcome measurements to be 'low risk' for these drugs.
The risk of bias for ITS studies can be evaluated using seven standard criteria but we found no ITS studies.
We included all the 'Risk of bias' criteria in the data abstraction form and independently scored criteria as 'yes' (adequate), 'no' (inadequate) or 'unclear'. We resolved disagreements by discussion and, where necessary, with a third review author. The risk of bias of included studies is summarized in the text and presented in the 'Risk of bias' section within the Characteristics of included studies table.
Two review authors (EC, PD) assessed the quality of evidence for each main outcome ‐ that is the extent of confidence in the estimate of effect across studies (high, moderate, low or very low) ‐ using the GRADE approach (Guyatt 2008).
Measures of treatment effect
For dichotomous variables, we used the risk ratio (RR).
When the outcomes were continuous variables, we calculated standardized mean differences (SMD) with 95% confidence intervals (CI). The SMD is a statistical measure of the impact of the intervention, which is independent of the units used to measure study outcomes. This measure allows studies of the same intervention using different outcomes to be compared. For example, measurement of drug concentrations in blood in different studies may use different assays in several laboratories and results may be reported in different units. The SMD compares differences between experimental and control groups to the standard deviation of the outcome for each study. Hence, a quantitative approximation can be made of the overall effect of decision support on plasma levels. Because SMD can be difficult to interpret (as it is reported in units of standard deviation), we also presented mean difference (MD), that is the absolute difference between the mean value in two groups, in relevant cases (when measurements were made on the same scale).
Clinical adverse events were expressed as RRs or rate ratios. In a randomized trial, rate ratios may often be very similar to RRs obtained after dichotomizing the participants, since the mean period of follow‐up should be similar in all intervention groups. Rate ratios and RRs will differ, however, if an intervention affects the likelihood of some participants experiencing multiple events.
The effect sizes were combined to give an overall effect for each subgroup of studies, using a random‐effects model. The random‐effects model was chosen because it does not assume that all interventions have the same underlying effect.
Unit of analysis issues
Analyses of studies using cluster randomization that do not account for the design effect risk inflating the type 1 error‐rate resulting in artificially narrow CIs (Ukoumunne 1999). We have reported potential errors and did not attempt to reanalyze data unless standard errors were correctly stated with number of clusters allowing the calculation of appropriate CIs.
Dealing with missing data
When the mean and standard deviation were missing, we estimated the mean from the median and standard deviation from the interquartile range or range (Hozo 2005). Missing data on outcomes or estimated data are explicitly indicated in the tables and text. We did not attempt to impute or model other missing data.
Assessment of heterogeneity
We explored heterogeneity using tests of heterogeneity and examinations of direction, magnitude and variability of effects. The statistical test for heterogeneity (Chi2 test) tests the null hypothesis that all studies in a meta‐analysis have the same underlying magnitude of effects. The I2 statistic quantifies the proportion of the variation in point estimates due to among‐study differences but is influenced by sample size. An I2 statistic greater than 50% may be considered as substantial heterogeneity and I2 statistic greater than 75% as considerable heterogeneity (Deeks 2011). High P value for the test of heterogeneity (P value > 0.1) and low I2 values do not necessarily indicate low heterogeneity (Guyatt 2011). Thus, we also manually examined the variability in point estimates across studies and the overlap of CIs.
Assessment of reporting biases
To be meaningful and appropriate, funnel‐plot asymmetry tests must be performed when four criteria are met: no significant heterogeneity (P value for the Chi2 test of heterogeneity > 0.10), low I2 statistic (< 50%), 10 or more studies with at least one with significant results, and a ratio of the maximal to minimal variance across studies greater than four (Ioannidis 2007). We could not assess publication bias because these conditions were not met.
Data synthesis
We grouped the results from the included studies by drug used and the effect aimed at for aminoglycoside antibiotics, amitriptyline, anaesthetics, insulin, anticoagulants, ovarian stimulation, anti‐rejection drugs and theophylline. We further grouped studies by type of outcome when appropriate (i.e. no evidence of heterogeneity).
We constructed Forest plots for the main outcomes without potential unit of analysis error for which data were available for more than one comparison.
The doses of drugs administered to participants and the number of dose changes per participant were described but not compared.
Subgroup analysis and investigation of heterogeneity
We considered the following potential sources of heterogeneity to explain variation in the results of the included studies:
-
time of delivery of advice (real time/delayed feedback);
-
location of advice (integrated in CPOE systems/other);
-
initiation of the computer advice (system‐initiated/user‐initiated);
-
advice given directly to the provider (direct intervention) versus the intermediate of a third party actor (indirect intervention);
-
type of hospital (outpatient/inpatient);
-
type of advice (recommendation/calculated dose proposed without possibility of change).
Sensitivity analysis
We had planned to perform a sensitivity analysis, excluding high risk of bias studies, but, since most of the identified studies had high risk of bias, we were unable to perform this analysis.
Results
Description of studies
See: Characteristics of included studies; Characteristics of excluded studies; Characteristics of ongoing studies.
Results of the search
The initial review identified 15 comparisons (14 articles) (Walton 2001), whereas the previous update of this review identified 26 comparisons (23 articles) that met inclusion criteria (Durieux 2008).
Figure 1 shows the study PRISMA flow chart (Moher 2009). We screened 5328 non‐duplicate records, of which 64 were identified through other sources than database searches (references from previous versions of this review and published before 1996, references from bibliography or website of screened articles or systematic reviews). We assessed 199 articles for potential inclusion: 136 articles identified from the new updated search strategy (from January 1996) (including eight from the previous version of the review) and 48 additional relevant articles from handsearches and reference lists of trials and systematic reviews, leading to 176 potential 'new' inclusions as compared with the previous version of the review; and 15 articles from the previous review and published before 1996.
Study flow diagram.
Since the search strategy missed studies we will search pre‐1996 for the next update of this review.
Included studies
This update identified 20 new trials with 21 comparisons, for a total of 42 trials and 46 comparisons and are included as follows: 25 comparisons (22 trials) from the previous update of this review (Durieux 2008), and 21 comparisons (20 trials) from this update.
These 42 trials were reported in 53 references (many reports for some trials). Three articles were separate cost‐effectiveness analyses: Jowett 2009 (main results in Poller 2008 PARMA 5; Poller 2009 DAWN AC), Rousseau 2010 (main results in Le Meur 2007), and Claes 2006 (main results in Claes 2005). One article was a separate safety analysis: Mihajlovic 2010 (main results in Mihajlovic 2003).
In four trials, two different comparisons were analyzed in one article, midazolam and fentanyl continuous infusion anaesthesia for cardiac surgery were independently titrated to maintain haemodynamic stability infusions but analyzed in the same population (Theil 1993 fentanyl; Theil 1993 midazolam). Therefore, these two drugs were reviewed separately. In one article, Vadher 1997 pop1 considered people starting warfarin with a targeted international normalized ratio (INR) between 2 and 3; Vadher 1997 pop2 considered people on long‐term treatment with a targeted INR between 3 and 4.5. Poller 1998 pop1 considered inpatients needing anticoagulant therapy (stabilized patients); Poller 1998 pop2 considered outpatients needing anticoagulant therapy (in the stabilization period). The results from two computer‐assisted dosage programs (DAWN AC and PARMA 5) were first published in one clinical endpoint report from the European Action on Anticoagulation (EAA), which gave the combined results (Poller 2008) and secondly published in two separate reports (Poller 2008 PARMA 5; Poller 2009 DAWN AC), because the computer‐assistance differed considerably between the participant centres in the study using the two alternative programs. We decided to include the two computer‐assisted dosage programs as two subgroup analyses.
All studies were RCTs, except two (Chertow 2001; Manotti 2001), which we classified as NRCTs.
In one publication concerning warfarin dosage adjustment (Carter 1987), three groups were studied: we reviewed only the comparison between the group using an analogue‐computer method and the group using empiric dosing (control). The third group, using a linear regression model, was excluded because it did not involve any computer assistance.
In one publication (Manotti 2001), two different groups of people were studied: one group starting oral anticoagulants (induction) and one group on long‐term treatment (maintenance). The maintenance study was not reviewed because of the absence of relevant data for the primary outcome.
In one publication, two different studies were reported: the first study considered people starting warfarin with a targeted INR between 2 and 3 (Vadher 1997 pop1); the second study considered people on long‐term treatment with a targeted INR between 3 and 4.5 (Vadher 1997 pop2).
In one publication (Fitzmaurice 2000), there were two levels of randomization. Practices were randomized to intervention or control. The study used two control populations: people individually randomly allocated to control in the intervention practices (intrapractice controls) and all participants in the control practices (interpractice controls). We did not analyze interpractice controls to avoid a possible unit of analysis error.
In one publication concerning oral anticoagulation therapy at steady state where randomization was at the GP practice level (reported in Claes 2005; Claes 2006), four groups were studied: we reviewed only the comparison between group A (Grol's multifaceted education: summary of the guidelines printed on the cover of a folder containing the anticoagulation files; information booklets on anticoagulation for their patients; website with guidelines, study design, and general information; newsletter sent every two months to inform the GPs on the study progress and requested them to send the anticoagulation files for checking) and group D (Grol's multifaceted education + DAWN AC computer‐assisted advice that generated a recommended dosing scheme and the time to next visit). Group A did not correspond to standard practice but it was considered as a control group because GPs in group D also received the multifaceted education. Groups B and C were excluded because the interventions were not computerized advice on drug dosage.
In one publication concerning insulin in cardiac surgery patients (Blaha 2009), three groups were studied: we reviewed only the comparison between the group using the Matias protocol based on the absolute glucose value and the group using computer‐based model predictive control algorithm with variable sampling rate (Enhanced software Model Predictive Control (eMPC)). The third group, using the Bath protocol based on the relative glucose change was excluded since most standard protocols in blood glucose management use the absolute glucose value and the Bath protocol had not been used in the hospital before the study.
Four comparisons/three trials included (Poller 1998 pop1; Anderson 2007; Poller 2008 PARMA 5; Poller 2009 DAWN AC) had duplicate publications (Poller 2002; Anderson 2008; Poller 2008), and three comparisons/two trials included (Le Meur 2007; Poller 2008 PARMA 5; Poller 2009 DAWN AC) were mentioned earlier in abstracts (Poller 2003; Le Meur 2007 extract).
Excluded studies
We excluded 143 of the 199 full‐text articles assessed for eligibility: 72 because the intervention was not a computerized drug dosage, 38 for an inappropriate design, 12 for absence of relevant data for primary outcome, seven because the patient aid was not under physician control, five because the dose advice was not individualized, three discussions, two conferences publications where contact to author failed (Ghazal‐Aswad 1997; Tomek 2011), one abstract published elsewhere (van Leeuwen 2005), one because some participants were already included in another publication (Jankovic 1999) and one comment (Ligtenberg 2006).
One study included in the previous update of the review was excluded because the intervention corresponded to a closed‐loop system (automatic optimization of the infusion rate of sodium nitroprusside achieved by an integrated hardware‐software closed‐loop controller implemented as a small bedside device) (Ruiz 1993).
Ongoing studies
A total of 17 studies are awaiting classification (possibly relevant ongoing studies published between January 2012 and January 2013). Three references corresponding to conference publications not published at the time of the search for this update (Overgaard 2010; Anderson 2011; Nieuwlaat 2011), have been published since then (Anderson 2012; Nieuwlaat 2012; Rasmussen 2012).
Characteristics of the providers
The providers were primarily doctors, although 14 studies (33.3%) targeted several categories of healthcare professionals including pharmacists (Carter 1987; White 1987; Destache 1990; Leehey 1993; Mungall 1994; Anderson 2007), nurses (Vadher 1997; Vadher 1997 pop2; Vadher 1997 pop1; Ageno 1998; Ageno 2000; Fitzmaurice 2000; Blaha 2009), or other healthcare professionals (Claes 2005; Claes 2006; Saager 2008). Three studies addressed only nurses' behaviour (White 1991; Pachler 2008; Cordingley 2009).
Twenty studies (47.6%) were conducted in North America (17 in the US, three in Canada) and 15 (35.7%) in Europe (three studies with numerous countries Poller 1998 pop1; Poller 1998 pop2; Plank 2006; Cordingley 2009). Two studies took place in New Zealand (Begg 1989; Hickling 1989), one in Australia (Hurley 1986), one in Israel (Verner 1992), one in Norway (Asberg 2010), and one in Serbia (Mihajlovic 2003; Mihajlovic 2010). One study was conducted in 13 countries from Europe, Israel and Australia (Poller 2008 PARMA 5; Poller 2009 DAWN AC).
Thirty‐one studies (73.8%) were conducted in one centre, and seven studies (16.7%) in two to five centres. One study took place in 11 centres (Le Meur 2007); one study included 12 practices (Fitzmaurice 2000); one study was conducted in 32 centres from Europe (29 centres), Israel (two centres), and Australia (one centre) (Poller 2008 PARMA 5; Poller 2009 DAWN AC); one study included 96 GPs regrouped in 66 GP practices (Claes 2005; Claes 2006).
Twenty‐seven comparisons (58.7%) included fewer than 100 participants in the analyses (median: 80 participants, mean: 779 participants).
Target behaviour
The target behaviour of the healthcare provider was the prescription and the dosing of drugs.
Characteristics of the interventions
Most of the studies provided advice about appropriate drug dosages to healthcare professionals who then decided whether to follow this or not. Fifteen studies (35.7%) (18 comparisons) evaluated anticoagulants, 10 studies (23.8%) evaluated the administration of insulin, five studies (11.9%) evaluated the administration of aminoglycoside antibiotics, four studies (9.5%) evaluated theophylline, four studies (9.5%) evaluated anti‐rejection drugs requiring adjustments for renal impairment, two studies (4.8%) (three comparisons) evaluated computer‐controlled infusions of anaesthetic agents, one study (2.4%) evaluated amitriptyline in the treatment of major depressive episodes and one study (2.4%) evaluated ovarian stimulation by gonadotropins.
Most of the computer support systems used a mathematical model of the pharmacokinetics of the drug to predict the required dose. These models represent the compartments in the body in which the drug is distributed, with rate constants determining the movement of the drug between different compartments. These systems allowed the operator to specify a target serum drug level, which the computer attempted to achieve using Bayesian forecasting methods. Where the effect of the drug was more important than the serum level, pharmacodynamic parameters based on population data could be added to the model (White 1987).
Anticoagulants (fifteen studies, eighteen comparisons)
Fourteen studies evaluated oral anticoagulant and one study evaluated heparin (Mungall 1994). Five studies analyzed initiation of warfarin (Carter 1987; White 1987; Ageno 2000; Manotti 2001; Anderson 2007), with varying target INR ranges (see Characteristics of included studies). Four studies (five comparisons) analyzed long‐term warfarin therapy (White 1991; Vadher 1997 pop1; Ageno 2000; Fitzmaurice 2000; Claes 2005; Claes 2006), with varying target INR ranges (see Characteristics of included studies). One study analyzed heparin therapy (Mungall 1994). Six studies (five comparisons) analyzed both initiation and long‐term warfarin therapy with at least three months of follow‐up (Vadher 1997; Vadher 1997 pop2; Poller 1998 pop1; Poller 1998 pop2; Mitra 2005; Poller 2008 PARMA 5; Poller 2009 DAWN AC).
The setting was outpatient care for six studies (White 1991; Vadher 1997 pop1; Vadher 1997 pop2; Ageno 1998; Poller 1998 pop1; Poller 1998 pop2; Manotti 2001; Poller 2008), community‐based care in two studies (Fitzmaurice 2000; Claes 2005; Claes 2006), and inpatient care for six studies (Carter 1987: White 1987; Mungall 1994; Ageno 2000; Mitra 2005; Anderson 2007); it was mixed in one study (Vadher 1997).
The computer support systems were programs that were not integrated into a CPOE. The computer‐generated program DAWN AC (4S Information Systems Ltd.) was used in five studies (Poller 1998 pop1; Poller 1998 pop2; Ageno 2000; Claes 2005; Mitra 2005; Claes 2006; Poller 2009 DAWN AC). Two modules existed in the DAWN AC program (induction and maintenance). Four studies used a (Bayesian) computer pharmacokinetic or a pharmacodynamic model (Carter 1987; White 1987; White 1991; Mungall 1994), or both, whereas one used a pharmacogenetics model (Anderson 2007). One study used the PARMA (Program for Archive, Refertation, and Monitoring of Anticoagulated patients) software program developed in Italy (Manotti 2001), and one comparison used PARMA 5, a new version of the program (Poller 2008 PARMA 5). Other studies used dosage algorithms or prediction rules.
The advice was given in real time to the healthcare professional in all studies except three, where it was unclear (Vadher 1997; Poller 1998 pop1; Poller 1998 pop2; Claes 2005; Claes 2006). The computerized advice was user‐initiated in seven studies (White 1987; White 1991; Vadher 1997 pop1; Vadher 1997 pop2; Ageno 1998; Fitzmaurice 2000; Claes 2005; Claes 2006; Anderson 2007), system‐initiated in two studies (Manotti 2001; Poller 2008 PARMA 5; Poller 2009 DAWN AC), and it was unclear in six studies (Carter 1987; Mungall 1994; Vadher 1997; Poller 1998 pop1; Poller 1998 pop2; Ageno 1998; Mitra 2005). The intervention was direct in seven studies (White 1991; Vadher 1997 pop1; Vadher 1997 pop2; Ageno 1998; Fitzmaurice 2000; Ageno 2000; Manotti 2001; Anderson 2007), indirect in three studies (dosage determined by pharmacy in Carter 1987 and Mungall 1994, and the pathologist reviewed the computer‐generated advice in Claes 2005) and it was unclear in five studies (White 1987; Vadher 1997; Poller 1998 pop1; Poller 1998 pop2; Mitra 2005; Poller 2008).
Three comparisons provided the warfarin maintenance doses per participant (Carter 1987; Vadher 1997 pop1; Vadher 1997 pop2), and there was no significant difference between groups (MD ‐0.33 mg/day, 95% CI ‐1.18 to 0.53). One study reported significantly larger amounts of drug prescribed for people with high INR target both of warfarin (computer group = 33.3 mg/week versus manual = 31.3 mg/week; P value < 0.001), and acenocoumarol (computer group = 19.2 mg/week versus manual = 17.8 mg/week; P value < 0.01), whereas there was no significant difference between the doses prescribed for people with low INR target (Manotti 2001). One study analyzed the number of dose adjustments (Anderson 2007): pharmacogenetic guidance significantly decreased the number of required dose adjustments (by 0.62 dose adjustments per participant; 95% CI 0.04 to 1.19; P value = 0.035).
There were two eligible comparisons with potential unit of analysis error that analyzed the proportion of dose adjustments. In one comparison (Ageno 1998), the percentage of dose adjustments performed by the healthcare professionals was 47.4%, whereas the computer needed 31.3% dose adjustments, a statistically significant 34.0% relative reduction (95% CI ‐41.9% to ‐24.7%). In the other comparison (Ageno 2000), the proportion of dose adjustments was 48% in the computer group and 45% in the manual group, a non‐significant relative increase of 7% (95% CI ‐10% to 27%).
In Claes 2005, there was no significant difference among the groups in number of tests per participant per month and per cent of participants with treatment changes.
For heparin (Mungall 1994), the mean dose was not significantly different between the computer and the standard groups (MD 100 units/hour, 95% CI ‐96 to 296).
Insulin (ten studies)
Eight studies (80%) evaluated insulin in people admitted into the intensive care unit with hyperglycaemia: six studies in cardiac surgery patients (Plank 2006; Hovorka 2007; Kremen 2007; Saager 2008; Blaha 2009; Sato 2011) and two studies in critically ill patients (Pachler 2008; Cordingley 2009). One study was conducted in general medical inpatients with type 2 diabetes (Wexler 2010) and one study in diabetic outpatients (Augstein 2007).
The computer support system was integrated into a CPOE in one study (Wexler 2010), whereas other systems were software developed by companies. The software Model Predictive Control (MPC) was used in two studies (Plank 2006; Kremen 2007). The laptop‐based algorithm MPC is a model representing the glucoregulatory system. Glucose concentration, insulin dosage and carbohydrate intake are the input variables for the MPC. The insulin infusion rate is the output parameter based on hourly glucose sampling. eMPC was used in four studies (Hovorka 2007; Pachler 2008; Blaha 2009; Cordingley 2009). The eMPC is an enhanced version of the model predictive control algorithm (MPC), which additionally generates the time of the next glucose measurement with an interval between samples varying from 0.5 to 4 hours. All six studies were part of the CLINICIP (Closed Loop Insulin Infusion for Critically Ill Patients; www.clinicip.org), an integrated project funded by the European Community working towards the development of a closed‐loop system to achieve safe tight glucose control in intensive care patients. Three studies used other software that take account of the characteristics of glucose dynamics (KArlsburg Diabetes Management System (KADIS): Augstein 2007, EndoTool Glucose Management System (MD Scientific): Saager 2008, GIN Computer Software (GINCS): Sato 2011), and one study used a weight‐based insulin dose calculator (Wexler 2010).
The advice was given in real time to the healthcare professional in all studies except one, where it was unclear (Augstein 2007). The computerized advice was user‐initiated in six studies (Plank 2006; Hovorka 2007; Kremen 2007; Augstein 2007; Pachler 2008; Blaha 2009), system‐initiated in two studies (Saager 2008; Wexler 2010), and it was unclear in two studies (Cordingley 2009; Sato 2011). The intervention was direct in seven studies (Hovorka 2007; Pachler 2008; Saager 2008; Cordingley 2009; Blaha 2009; Wexler 2010; Sato 2011), indirect in one study (eMPC algorithm was run by study personnel (input glucose and change of Insulin infusion rate) under the supervision of the healthcare professional in Plank 2006) and it was unclear in two studies (Kremen 2007; Augstein 2007).
Seven comparisons provided the insulin doses. The statistical heterogeneity was moderate. In one study (Hovorka 2007), the mean insulin infusion rate was significantly higher in the computerized glucose management group than standard management protocol group (MD +2.10 insulin units/hour, 95% CI 0.77 to 3.43), whereas there were no significant differences in the others (four comparisons with higher doses in the computer group, two comparisons with higher doses in the standard group). Overall, the insulin doses were higher in the computer groups, but this was not significantly different (pooled SMD 0.23, 95% CI ‐0.04 to 0.50).
One study reported no differences in the amount of glucose administered between the computer group (79.4 ± 24 g) and the manual group (81.6 ± 28 g) during the study period (before, during and after cardiopulmonary bypass) (MD ‐2.2 g, 95% CI ‐19.2 to 14.8) (Sato 2011).
One study analyzed the number of times the insulin rate was changed in 72 hours: the insulin infusion rate was altered a mean of 23.5 times more (95% CI 19.0 to 28.0) in the computer group than in the control group (35.5, 95% CI 31.1 to 39.9 versus 12.0, 95% CI 10.3 to 13.7) (Pachler 2008).
Aminoglycoside antibiotics (five studies)
Five studies (Begg 1989; Hickling 1989; Destache 1990; Burton 1991; Leehey 1993) evaluated the administration of aminoglycoside in inpatient care. All computer support systems used a (Bayesian) pharmacokinetic model with the advice given in real time to the healthcare professional, not integrated into a CPOE, and it was not clear if the computerized advice was user‐initiated. The intervention was indirect in two studies (Destache 1990: a clinical pharmacokinetic service reviewed the initial aminoglycoside dose and dosing interval and made an oral recommendation to the attending physician or resident; Leehey 1993: the orders for aminoglycoside dosing were written by a pharmacist with countersignature by a physician and the duration of antibiotic therapy as well as other aspects of clinical care were determined by the primary physicians).
One comparison provided outcomes for the analysis on initial and maintenance doses (Burton 1991). There was no statistical difference between groups for initial doses (MD +8 mg/day, 95% CI ‐11 to 27) or for maintenance doses (MD +11 mg/day, 95% CI ‐16 to 38).
Two studies reported data on total administered dose with different magnitude of effects. In one study (Begg 1989), the aminoglycoside dose per day was significantly higher in the pharmacokinetic group than in the standard group (+109 g, 95% CI 67 to 151). In the other study (Leehey 1993), the milligrams per dose were higher and number of doses per day was lower in the pharmacist‐directed dosing group compared with the standard group (milligrams/dose: 107 ± 21 versus 91 ± 26; doses/day: 2.0 ± 0.6 versus 2.3 ± 0.5) but the mean total doses of aminoglycoside were not significantly different between groups (pooled MD 141 mg, 95% CI ‐342 to 624).
One study analyzed the number of dosage changes and indicated higher dosage changes in the group with pharmacokinetic service recommendation than in the control group (MD +0.50 dosage change, 95% CI 0.21 to 0.79) (Destache 1990).
Theophylline (four studies)
Four studies evaluated theophylline (Hurley 1986; Gonzalez 1989; Verner 1992; Casner 1993), a drug that is not considered as the first choice of treatment of asthma at present. However, monitoring serum concentrations of theophylline is essential to ensure that non‐toxic doses are achieved (National Asthma 2002). There were no studies on recently introduced drugs where it is considered important to monitor drug levels such as for glycopeptides, antifungal (fluconazole) and antiretroviral drugs.
The setting was inpatient care for all four studies.
All computer support systems used a Bayesian compartmental pharmacokinetic model. The advice was given in real time to the healthcare professional in three studies, whereas it was unclear in one study (Verner 1992). The computer support system was integrated into a CPOE with a direct intervention in one study (Casner 1993), whereas it was unclear for the others. The computerized advice was user‐initiated in two studies and it was unclear in two studies.
Three comparisons provided data on initial dose with substantial heterogeneity (Hurley 1986; Gonzalez 1989; Verner 1992). The theophylline initial dose was significantly higher in the computer group (SMD 1.7, 95% CI 0.7 to 2.6), whereas the difference did not reach significance in the two other studies (SMD 0.2, 95% CI ‐0.1 to 0.6). Two comparisons provided data on maintenance dose and indicated higher doses in the computer group (SMD 0.8, 95% CI 0.5 to 1.1) (Hurley 1986; Gonzalez 1989).
Anti‐rejection drugs (four studies)
Four studies evaluated anti‐rejection drugs requiring adjustments for renal impairment (Chertow 2001; Le Meur 2007; Asberg 2010; Terrell 2010). Two studies evaluated high‐use medications that required adjustments for renal impairment (Chertow 2001; Terrell 2010), one study evaluated cyclosporine A (CsA) in the early post‐transplant phase (Asberg 2010), and one study evaluated mycophenolate mofetil (MMF) dosing in renal transplant patients (Le Meur 2007).
The studies evaluating medications that required adjustments for renal impairment took place in inpatient care, with the computer support system integrated into a CPOE, and the advice system initiated and given in real time. In the study evaluating CsA in early post‐transplant phase (Asberg 2010), the individual computer dosing of CsA doses were calculated by a population pharmacokinetic model and suggested to the physician (patients were admitted in nephrology and had a standard clinical follow‐up). In the study evaluating MMF dosing in renal transplant patients (Le Meur 2007), the MMF dose adjustments in the concentration‐controlled regimen were calculated by a computer program to reach a mycophenolic acid (MPA) area under the curve (AUC) target of 40 mg.h/L and were proposed to the physician.
No study reported data on drug dosages. One eligible study with potential unit of analysis error showed that a computerized decision support system for prescribing drugs in people with renal insufficiency improved the proportion of appropriate orders (RR 1.71, 95% CI 1.64 to 1.78) (Chertow 2001).
Anaesthetic agents (two studies, three comparisons)
Two studies (three comparisons) evaluated computer‐controlled infusions of anaesthetic agents (Rodman 1984; Theil 1993 fentanyl; Theil 1993 midazolam). One study evaluated the lidocaine therapy (Rodman 1984), whereas one study evaluated fentanyl and midazolam infusions (Theil 1993 fentanyl; Theil 1993 midazolam).
The setting was inpatient care for the two studies. The computer support system for initial therapy of lidocaine was an individualized linear two‐compartment pharmacokinetic model not integrated into a CPOE, and the advice was user‐initiated and given at real time. The computer‐controlled pump for fentanyl and midazolam infusions used a pharmacokinetic model integrated into a CPOE, and the advice was system‐initiated.
Rodman 1984 provided outcomes for the initial, maintenance and total doses. Computerized advice had no significant effect on lidocaine dosage (SMD 2.5, 95% CI 1.3 to 3.8 for initial dose; ‐0.2, 95% CI ‐1.0 to 0.7 for maintenance dose; 0.2, 95% CI ‐0.7 to 1.1 for total dose).
Theil 1993 fentanyl and Theil 1993 midazolam provided outcomes for the initial, maintenance and total doses for both fentanyl and midazolam infusions. Computerized advice had no significant effect on fentanyl drug dosage (SMD 0.5, 95% CI ‐0.3 to 1.3 for initial dose; 0.0, 95% CI ‐0.8 to 0.8 for maintenance dose; 0.5, 95% CI ‐0.3 to 1.3 for total dose), but reduced significantly midazolam initial, maintenance and total drug doses (SMD ‐1.9, 95% CI ‐2.9 to ‐0.9 for initial dose; ‐1.2, 95% CI ‐2.1 to ‐0.3 for maintenance dose; ‐1.1, 95% CI ‐1.9 to ‐0.2 for total dose). There was no significant difference in the number of infusion changes during the cardio‐pulmonary bypass (MD ‐0.2, 95% CI ‐1.0 to 0.6 for fentanyl; 0.5, 95% CI ‐0.3 to 1.3 for midazolam).
Antidepressants (one study)
One study evaluated amitriptyline in the treatment of major depressive episodes (Mihajlovic 2003; Mihajlovic 2010). The study took place in one psychiatric clinic of a clinical hospital centre. The computer‐aided dose of amitriptyline was calculated using the modified Bayesian method. It was not clear if the computer support system was integrated into a CPOE, if the intervention was direct or if the advice was user‐initiated and given at real time.
The drug daily doses of amitriptyline plus nortriptyline at day 14 were significantly lower when they were individualized compared with empiric doses (133.3 mg, 95% CI 126.7 to 140.0 versus 148.3 mg, 95% CI 140.2 to 156.4).
Gonadotropins (one study)
One study evaluated ovarian stimulation by gonadotropins (Lesourd 2002). The study included women from three centres who were undergoing ovarian stimulation to treat infertility. The software (GonaSoft) was created by the author to help clinicians to monitor ovarian stimulation and to provide a tool for evaluation of efficiency and complications. The software was not integrated into a CPOE. The intervention was direct with the advice given at real time.
There was no significant difference between groups in the number of follicle‐stimulating hormone units administered (860 units, 95% CI 776 to 944 in the intervention group versus 938 units, 95% CI 825 to 1051 in the control group).
Risk of bias in included studies
Risk of bias graph: review authors' judgements about each risk of bias item presented as percentages across all included studies.
Risk of bias summary: review authors' judgements about each risk of bias item for each included study.
Four studies were not assessed for the risk of bias, since they were cost‐effectiveness or safety analyses conducted as a part of previous included studies (same trial): Claes 2006 (see Claes 2005); Jowett 2009 (see Poller 2008 PARMA 5; Poller 2009 DAWN AC); Mihajlovic 2010 (see Mihajlovic 2003); and Rousseau 2010 (see Le Meur 2007).
Random sequence generation (selection bias)
The sequence generation process used a random component in 21 studies (50%). A non‐random method was used in two studies (5%): in one study the study periods were four alternating eight‐week blocks of intervention and control subperiods (Chertow 2001); and in one study, the randomization was based on the final digit in the patient's identification card number (odds versus even) (Verner 1992). The sequence generation process was not specified in 19 studies (45%).
Allocation concealment (selection bias)
The allocation was adequately concealed in 16 studies (38%). The unit of allocation was institution, team or provider in five studies (Burton 1991; Fitzmaurice 2000; Claes 2005; Terrell 2010; Wexler 2010), but cluster effect was not taken into account in statistical analysis for Burton 1991 and Wexler 2010. The unit of allocation was by participant or episode of care and there was some form of centralized randomization scheme in 11 studies (on‐site computer system: eight studies, sealed opaque envelopes: three studies). The allocation was not considered adequately concealed in the two studies (5%) where a non‐random method was used (Verner 1992; Chertow 2001). The allocation concealment was not specified in 24 studies (57%).
Baseline outcome measurements similar
The mean blood glucose was measured prior to the intervention and no important differences were present across study groups in eight (80%) of the 10 studies on insulin drug administration. One study reported the blood glucose at entry for each centre but not across study group (Plank 2006), and in one study there were many data reported on blood glucose but it was unclear if results given were baseline measurements (Wexler 2010).
Other studies were considered at 'low risk' because baseline outcome measurements were not relevant (baseline data realized before medication intake).
Baseline characteristics similar
Of the five studies where the unit of allocation was by institution, team or professional, only one reported the characteristics of providers (Terrell 2010). Baseline characteristics of participants were reported and similar in 26 studies (62%). Seven studies (17%) reported differences between participants in control and intervention groups, one study reported the participants' characteristics only in a subgroup of participants (Carter 1987), one study reported the participants' characteristics by centres (Plank 2006), and three studies did not report participants' characteristics (Hickling 1989; Poller 1998 pop1; Poller 1998 pop2; Claes 2005). Four studies mentioned participants' characteristics in text but no data were presented (Gonzalez 1989; Ageno 2000; Chertow 2001; Asberg 2010).
Incomplete outcome data (attrition bias)
In 18 studies (43%), the missing outcome measures were unlikely to bias the results (e.g. the proportions of missing data were similar in the intervention and control groups or the proportions of missing data were unlikely to overturn the study result). In 13 studies (31%), missing outcome measures were not specified in the paper or it was unclear if missing data could overturn the study result. In 11 studies, the missing outcome data were likely to bias the results (Carter 1987; Begg 1989; Gonzalez 1989; Hickling 1989; Destache 1990; Burton 1991; Casner 1993; Vadher 1997; Fitzmaurice 2000; Augstein 2007; Terrell 2010).
Blinding of outcome assessment (detection bias)
In 36 studies (86%), the assessment of primary outcome was objective or blinded. In six studies (14%), there was a risk of detection bias: five studies used clinical events as main outcomes and they were not clearly defined (Gonzalez 1989; Ageno 2000; Fitzmaurice 2000; Mitra 2005; Le Meur 2007); and one study used a hetero‐questionnaire as main outcome (Mihajlovic 2003).
Blinding of participants and personnel (performance bias)
When studies were randomized by participant, the same healthcare professional may have given treatment both to intervention and control groups: it is possible that computerized advice influenced the treatment of the control groups. Protection against contamination was considered to be done only in three studies (Theil 1993 fentanyl; Theil 1993 midazolam; Claes 2005; Claes 2006; Anderson 2007), and it was unclear for three studies were it was possible that communication between intervention and control professionals could have occurred (Fitzmaurice 2000; Terrell 2010; Wexler 2010).
Thirty‐seven studies (88%) randomized the participants, but in two of them the participants were blinded (Theil 1993 fentanyl; Theil 1993 midazolam; Anderson 2007). Two studies randomized the GP practices (Fitzmaurice 2000; Claes 2005; Claes 2006), one study randomized 42 emergency medicine faculty and resident physicians from one urban public hospital (Terrell 2010), and one study randomized seven teams of providers (42 internal medicine residents) in general medical acute care units of one medical centre (Wexler 2010). There was a risk of performance bias in one study where the house staff teams were randomized (17 house staff teams in one Veterans Administration medical centre) and at the end of each four months during the study, intervention groups were changed to control and vice versa (Burton 1991).
Selective reporting (reporting bias)
The relevant outcomes were reported in the results section in 39 studies (93%). In three studies, the outcomes were not presented in the methods section (Rodman 1984; White 1991; Kremen 2007).
Other bias
There was no evidence of other risk of biases in 34 studies (81%), an evidence of other risk of bias in three studies (7%) and it was unclear in five studies (12%).
In one study, an erratum had been published because there were some inconsistencies in the text and tables; we found other inconsistencies in tables, full text, and abstract; the author was contacted but had not replied (Cordingley 2009, by January 2012). There is a risk of selection bias for one study (Mihajlovic 2003; Mihajlovic 2010): there was a first publication in 1999 including 15 participants during 1997 (Jankovic 1999, study excluded because the author confirmed that the participants were included in the Mihajlovic study), a second publication in 2003 on main outcomes including 60 participants during 1997 (Mihajlovic 2003), and a third publication in 2010 on clinical adverse events (Mihajlovic 2010). In one study, there was a risk of contamination due to logistical problems ("it was difficult to shield the clinicians from the CDSS suggestions") and the nurse practitioners used the computer‐decision support system and were compared with the clinician group of three junior doctors undergoing general professional training in general medicine (Vadher 1997 pop1; Vadher 1997 pop2).
There was potential other risk of bias in five studies: in one study, the Cockcroft‐Gault formula might overestimate renal function when the serum creatinine was increasing, and underestimate renal function when the serum creatinine was decreasing (Chertow 2001); one study was described as a pilot study (10 participants in each group), the article was in Czech language so it was difficult to understand all details (authors were contacted in February 2012 but queries remained unanswered) (Kremen 2007); in one study, few results were reported (one sentence in the results) so there was not enough information to evaluate the bias of the study (Lesourd 2002); in one study, only 20 participants were included (Rodman 1984); and in one study, 46% participant visits were excluded (only prescription that required dosage adjustment were analyzed), there was no adjustment for within‐patient correlation, and providers in the intervention group initially prescribed targeted medications more often than control physicians did and consequently had substantially more opportunities to adjust dosing (Terrell 2010).
Protection against bias
No study met the nine previous criteria for protection against bias. Two studies met eight criteria (Anderson 2007; Sato 2011), nine studies met seven criteria (Begg 1989; Destache 1990; Burton 1991; Casner 1993; Theil 1993 fentanyl; Theil 1993 midazolam; Claes 2005; Claes 2006; Pachler 2008; Wexler 2010), 16 studies met six criteria (Hurley 1986; White 1987; Hickling 1989; White 1991; Mungall 1994; Vadher 1997; Poller 1998 pop1; Poller 1998 pop2; Fitzmaurice 2000; Augstein 2007; Hovorka 2007; Le Meur 2007; Poller 2008 PARMA 5; Blaha 2009; Poller 2009 DAWN AC; Terrell 2010), nine studies met five criteria (Gonzalez 1989; Verner 1992; Leehey 1993; Vadher 1997 pop1; Vadher 1997 pop2; Manotti 2001; Saager 2008; Cordingley 2009; Asberg 2010), seven studies met four criteria (Carter 1987; Ageno 1998; Ageno 2000; Chertow 2001; Mihajlovic 2003; Mitra 2005; Plank 2006; Mihajlovic 2010), and three studies met three criteria (Rodman 1984; Lesourd 2002; Kremen 2007).
Power calculation
Fifteen studies (36%) reported a sample size calculation, among them, one study was presented as a non‐inferiority trial but no margin was addressed (Pachler 2008). The power calculation was not reported in 22 studies (53%) and in five studies (12%) the authors specifically reported that the study might be underpowered (White 1987; Gonzalez 1989; Casner 1993; Mitra 2005; Asberg 2010).
Effects of interventions
See: Summary of findings for the main comparison Computerized advice on drug dosage for leading serum concentrations within therapeutic range; Summary of findings 2 Computerized advice on drug dosage for leading physiological parameters within therapeutic range; Summary of findings 3 Computerized advice on drug dosage for reducing time to achieve therapeutic control; Summary of findings 4 Computerized advice on drug dosage for leading to fewer clinical adverse events; Summary of findings 5 Saving healthcare resources for saving healthcare resources
See: summary of findings Table for the main comparison; summary of findings Table 2, summary of findings Table 3; summary of findings Table 4; summary of findings Table 5.
Hypothesis 1. Decisions on drug dosage based on computer advice lead more often to drug levels within the therapeutic range
For this comparison, the outcomes analyzed were: the serum concentrations (Analysis 1.1; Analysis 1.2; Figure 4; Figure 5), the proportion of time for which the plasma drug concentrations was within the therapeutic range (Analysis 1.3; Figure 6), the proportion of participants with plasma drug concentrations within the therapeutic range (at fixed time) and the proportion of participants with toxic drug levels (Analysis 1.4; Figure 7).
Forest plot of comparison: 1 Serum concentrations and therapeutic range, outcome: 1.1 Serum concentrations (mg/L) ‐ part A (SMD > 0 in favour of the intervention).
Forest plot of comparison: 1 Serum concentrations and therapeutic range, outcome: 1.2 Serum concentrations (ng/L) ‐ part B (SMD < 0 in favour of the intervention).
Forest plot of comparison: 1 Serum concentrations and therapeutic range, outcome: 1.3 Proportion of participants within therapeutic range.
Forest plot of comparison: 1 Serum concentrations and therapeutic range, outcome: 1.4 Proportion of participants with toxic drug levels.
Serum concentrations
There was a high clinical and statistical heterogeneity between drugs so we did not pool the results. Since the interpretation of the direction of change in serum concentration varied according to the drug, we grouped the drugs in two Forest plots (part A and part B). Part A (Analysis 1.1; Figure 4) includes the drugs for which an SMD greater than 0 corresponds to a difference in favour of the intervention; Part B (Analysis 1.2; Figure 5) includes the drugs for which an SMD less than 0 corresponds to a difference in favour of the intervention.
Anticoagulants (fifteen studies, eighteen comparisons)
No data available.
Insulin (ten studies)
No data available.
Aminoglycoside antibiotics (five studies)
Four comparisons (Begg 1989; Hickling 1989; Burton 1991; Leehey 1993) analyzed the aminoglycoside target peak concentration and the objective was to obtain a higher target peak concentration since target peak concentrations are often not met. Heterogeneity was moderate (inconsistency I2 = 51%, P value from the Chi2 = 0.11) but the CIs for the results of individual studies overlapped and all studies resulted in a significant higher aminoglycoside peak concentration in the computer group (pooled SMD 0.79, 95% CI 0.46 to 1.13), which was in favour of the intervention.
In Burton 1991, the proportion of participants with maximum peak serum aminoglycoside concentrations greater than 4 mg/L was greater in the Bayesian pharmacokinetic dosing group (RR 1.37, 95% CI 1.11 to 1.70; odds ratio (OR) 3.19, 95% CI 1.46 to 6.94).
One study analyzed the aminoglycoside trough concentration (residual) and the objective was to obtain a lower trough concentration (Leehey 1993): the mean serum trough drug concentration was lower in the pharmacist‐directed dosing group compared with the control group (SMD ‐0.42, 95% CI ‐0.74 to ‐0.09).
Theophylline (four studies)
Four comparisons analyzed the theophylline concentration: Hurley 1986 compared the mean serum concentration at day two, Gonzalez 1989 compared the concentration four hours post load, Verner 1992 compared the serum concentration 20 minutes after completion of loading dose infusion and Casner 1993 compared the serum level just before discontinuation of the infusion. The Forest plot and heterogeneity statistics showed high statistical heterogeneity (inconsistency I2 = 76%, P value from the Chi2 = 0.005). The theophylline concentration was significantly higher in the computer group in one comparison (Gonzalez 1989), which was in favour of the intervention; tended to be higher in the computer group (although it did not reach statistical significance) in two comparisons (Verner 1992; Casner 1993); and tended to be lower in one comparison (Hurley 1986). The pooled difference was not significant (SMD 0.41, 95% CI ‐0.20 to 1.02).
Anti‐rejection drugs (four studies)
No data available.
Anaesthetic agents (two studies, three comparisons)
Rodman 1984 showed that computer‐assisted initial lidocaine therapy significantly increased the lidocaine concentration (SMD 1.32, 95% CI 0.33 to 2.32), which is in favour of the intervention (Analysis part A).
In two comparisons dealing with anaesthesiology (Theil 1993 fentanyl; Theil 1993 midazolam), the objective was to obtain a lower administered drug dose in order to provide a reduction in time for extubation. Computerized advice had no effect on fentanyl serum concentrations but significantly reduced midazolam serum concentrations (SMD ‐1.43, 95% CI ‐2.34 to ‐0.51), which was in favour of the intervention (Analysis part B).
Antidepressants (one study)
The steady‐state plasma concentrations of amitriptyline plus nortriptyline during the treatment course (day 14) was significantly lower in the individualized regimen compared with the empiric dose regimen (SMD ‐0.68, 95% CI ‐41.20 to ‐0.16), which was in favour of the intervention.
Gonadotropins (one study)
No data available.
Proportion of time for which the plasma drug concentrations were within the therapeutic range/proportion of participants with plasma drug concentrations within the therapeutic range (at a fixed time) (Analysis 1.3)
Anticoagulants (fifteen studies, eighteen comparisons)
No data available.
Insulin (ten studies)
No data available.
Aminoglycoside antibiotics (five studies)
Several studies have suggested that the mortality in people with severe infections treated with aminoglycosides may be substantially reduced if adequate peak plasma concentrations are achieved early in the course of treatment (Moore 1984a, Moore 1984b, Moore 1987). Thus, accurate dose prescription is important, not only to avoid toxicity associated with overdosage, but more important to avoid the higher mortality associated with underdosage during the first one or two days of treatment.
Three studies analyzed the proportion of participants with aminoglycoside peak concentration adequate: one study (Destache 1990) considered the first peak aminoglycoside serum 'adequate' 30 minutes after infusion whereas two studies (Hickling 1989; Begg 1989) considered the peak plasma concentrations within 6 to 10 mg/L after two days, so we pooled only the results of these two last studies. There was no evidence of difference in Destache 1990 (RR 1.02, 95% CI 0.63 to 1.64), whereas there were significantly more participants within the therapeutic range after two days in the computer group for the two other studies (pooled RR 4.44, 95% CI 1.94 to 10.13).
Hickling 1989 and Begg 1989 also analyzed the proportion of participants with aminoglycoside peak and trough concentrations adequate: Hickling 1989 considered the peak trough concentrations lower than 2 mg/L after two days whereas Begg 1989 considered the peak trough concentrations within 1 to 2 mg/L after two days. Both comparisons tended to have more participants with aminoglycoside peak and trough concentrations adequate in the computer group and the pooled effect showed a significant difference between groups (pooled RR 3.88, 95% CI 1.04 to 14.44).
Theophylline (four studies)
Hurley 1986 reported that during oral therapy more monitored than control participants had trough concentrations in the therapeutic range (RR 1.60, 95% CI 1.00 to 2.55).
One study reported that serum theophylline concentrations during maintenance therapy in the computer group were maintained within the therapeutic range (10 to 20 µg/mL) longer than in the control group (77% versus 51%, P versus < 0.05), but the study included a small number of participants (n = 25) (Verner 1992).
Anti‐rejection drugs (four studies)
In Le Meur 2007, there was no significant difference between concentration‐controlled doses and fixed‐dose MMF on the proportion of participants within therapeutic range at day 14 (MPA AUC > 30 mg.h/L). The AUC is a model for determining MPA exposure. MPA is the active metabolite of the inactive prodrug MMF. However, by day 14, median MPA exposure was significantly higher in the concentration‐controlled group than in the fixed‐dose group (with a majority of participants in the concentration‐controlled group, but not in the fixed‐dose group, having met targeted levels), and at one month, the concentration‐controlled group again had significantly higher median MPA AUC, with more than 90% of participants achieving target levels.
In Asberg 2010, the overall percentage of whole‐blood concentrations of cyclosporine within the therapeutic window was not different between groups (SMD 0.23, 95% CI ‐0.39 to 0.85; MD 3.80%, 95% CI ‐6.25% to 13.85%).
Anaesthetic agents (two studies, three comparisons)
No data available.
Antidepressants (one study)
No data available.
Gonadotropins (one study)
No data available.
Proportion of participants with toxic drug levels (Analysis 1.4)
Anticoagulants (fifteen studies, eighteen comparisons)
No data available.
Insulin (ten studies)
No data available.
Aminoglycosides (five studies)
No data available.
Theophylline (four studies)
Two studies analyzed the proportion of participants with toxic drug levels. In Hurley 1986, fewer monitored participants in whom infusion rates were based on pharmacokinetic analysis had serum concentrations in the toxic range but the difference was not significant (RR 0.50, 95% CI 0.23 to 1.10; OR 0.38, 95% CI 0.13 to 1.09). Casner 1993 found no statistical difference between the empiric and kinetic groups in the number of toxic (> 20 mg/L) levels (one in each group).
Anti‐rejection drugs (four studies)
In Terrell 2010, where the physicians were randomized to decision support intervention group or control group, usual care physicians were more likely to dose medications excessively than intervention physicians were (OR 3.9, 95% CI 1.7 to 9.0, according to a mixed‐effects logistic regression to adjust for within‐physician correlation). After adjusting for participant age, sex, race and physician status, this difference remains significant in favour of intervention (OR usual care versus intervention 4.3, 95% CI 1.4 to 12.8).
Anaesthetic agents (two studies, three comparisons)
No data available.
Antidepressants (one study)
No data available.
Gonadotropins (one study)
No data available.
Summary
The results differed according to drugs. In summary, in spite of high heterogeneity and small number of studies, comparisons on serum concentrations were in favour of the computer group for aminoglycoside antibiotics (higher target peak and lower trough concentrations) and antidepressants (one study). Results were contrasted for theophylline and anaesthetic agents.
Computerized advice improved the proportion of participants for aminoglycoside antibiotics and theophylline (one study), and the proportion of time for theophylline (one study) for which the plasma drug concentrations were within the therapeutic range, but not for anti‐rejection drugs before day 14.
There were three comparisons on toxic drug levels: there was no evidence of difference for theophylline (two studies) whereas the CPOE with decision support significantly reduced excessive dosing of targeted medications in one study on people with renal impairment where physicians were randomized.
Hypothesis 2. Decisions on drug dosage based on computer advice lead more often to a physiological parameter being maintained within the desired range (e.g. blood pressure or prothrombin time)
For this comparison, the outcomes analyzed was the proportion of time for which the studied physiological parameter was maintained within the target range (Analysis 2.1; Analysis 2.2; Analysis 2.3; Figure 8; Figure 9; Figure 10).
Forest plot of comparison: 2 Physiological parameters, outcome: 2.1 Oral anticoagulants: % time in target INR range.
Forest plot of comparison: 2 Physiological parameters, outcome: 2.2 Insulin: % time in target glucose range.
Forest plot of comparison: 2 Physiological parameters, outcome: 2.3 Insulin: mean blood glucose (mg/dL).
Oral anticoagulants (fifteen studies, eighteen comparisons): proportion of time in target international normalized ratio range
The main outcome reported was the percentage of time spent in target INR range calculated for each participant as a mean time in range using interpolation methods between INR values. Six comparisons analyzed the mean % of TIR using various INR ranges (see Description of the intervention) (White 1987; Poller 1998 pop1; Poller 1998 pop2; Anderson 2007; Poller 2008 PARMA 5; Poller 2009 DAWN AC). The magnitude of effects differed between studies and the statistical heterogeneity was high (I2 = 79%) but all studies were in favour of the computer groups: the difference was significant in four studies (White 1987; Poller 1998 pop1; Poller 2008 PARMA 5; Poller 2009 DAWN AC), and did not reach significance in two studies (Poller 1998 pop2 considering outpatients in the stabilization period; and Anderson 2007 considering people starting oral anticoagulation). The pooled SMD favoured the computer group (SMD 0.19, 95% CI 0.06 to 0.33; MD 3.68%, 95% CI 0.90% to 6.45%) (Analysis 2.1; Figure 8).
In one study where GP practices were randomized, there was a significant increase in percentage of time within 0.5 INR from target, from 49.5% at baseline to 60% after implementing the different interventions, but there was no evidence that the increases from baseline were different between the four intervention groups (P value = 0.8) (Claes 2005). The increase was +8% (95% CI 2.0% to 13.5%) in the group with multifaceted education and +11% (95% CI 5.5% to 16.5%) in the group with multifaceted education plus DAWN AC computer‐assisted advice.
Another comparison on people within target final PT time showed no evidence of difference in anticoagulant control in people whose dose was determined by computer, compared with those who were treated by a nurse specialist (RR 0.87, 95% CI 0.47 to 1.61) (White 1991). Finally, one comparison analyzed the proportion of participants reaching a stable state of anticoagulation (three INR measurements within therapeutic range) and was in favour of the computer group compared with the control group (RR 1.46, 95% CI 1.07 to 2.00) (Manotti 2001).
Five comparisons reported the number of days per 100 patient‐days of treatment spent in the INR therapeutic range but the inconsistency across studies was very high (combined incidence rate ratio 1.10, 95% CI 1.00 to 1.20, inconsistency I2 = 91%) (Vadher 1997; Vadher 1997 pop1; Vadher 1997 pop2; Ageno 1998; Mitra 2005): in two studies, participants in the intervention group spent significantly more time in the therapeutic range (Vadher 1997 pop1; Mitra 2005), whereas there was no evidence of difference in three studies.
Two comparisons analyzed the proportion of INRs measurements within therapeutic range and were in favour of computerized advice of drug dosage (RR 1.10, 95% CI 1.02 to 1.19) (Ageno 1998; Fitzmaurice 2000). Ageno 1998 and Ageno 2000 compared the proportions of INRs above 5: there was no evidence of difference between the computer and standard groups. In Claes 2005, during the intervention period there was a significant difference in per cent of participants with a least one INR above 5 between the four intervention groups (P value = 0.009), but there was no evidence that the decreases from baseline were different between the four intervention groups (P = 0.28) and group A (multifaceted education) and group D (multifaceted education plus DAWN AC computer‐assisted advice) had the closest results. In Ageno 1998, the INRs below 2 (underanticoagulated patients) were more in the computer‐controlled group than in the manual group (RR 1.55, 95% CI 1.11 to 2.16) whereas in Claes 2005 there was no statistically significant difference from baseline in the decrease of per cent of participants with at least one INR less than 2 (P = 0.67).
Insulin (ten studies): percentage of time in target glucose range (four studies), hyperglycaemia index (nine studies)
Four studies of the CLINICIP project analyzed the percentage of time in target glucose range (Plank 2006; Hovorka 2007; Blaha 2009; Cordingley 2009). The magnitude of effects differed between studies and the statistical heterogeneity was high (I2 = 83%) but all studies were in favour of the computer groups: the difference was significant in three studies (Plank 2006; Hovorka 2007; Cordingley 2009), and did not reach significance in one study (Blaha 2009). The pooled SMD was in favour of the computer group (SMD 1.27, 95% CI 0.56 to 1.98; MD 22.18%, 95% CI 9.94% to 34.43%) (Analysis 2.2; Figure 9).
In one study with a potential unit of analysis error, the percentage of time in target range was significantly higher with the computer software program (SMD 1.91, 95% CI 1.70 to 2.11; MD 21.70%, 95% CI 20.02% to 23.38%) (Sato 2011). In the same study, participants in the computer group were significantly more likely to have all measurements within the target glycaemic range than participants in the manual group (RR 9.00, 95% CI 1.27 to 63.89).
In a CBA study, the change in A1C (primary outcome) was ‐0.34% ± 0.49% in the group with the diabetes management system‐based decision support compared with +0.27% ± 0.67% in the group without the system (MD ‐0.60%, 95% CI ‐0.96% to ‐0.25%) (Augstein 2007).
Two studies analyzed the hyperglycaemic index (HGI) for the assessment of glucose control (Pachler 2008; Cordingley 2009). In Pachler 2008, the HGI was defined as the AUC above the upper limit of normal (glucose level 6.1 mmol/L, modified from the original 6.0 mmol/L) divided by the total length of stay (time in study). The advantage of this measure of glucose control is the independence of the number of measurements, and it is not falsely lowered by hypoglycaemic values. In theory, the best HGI of 0.0 mM indicates that all glucose values were below the upper target limit. An HGI around 2.0 mM shows that the person was exposed on average to glucose values of 8.1 mM (exceeding the upper target limit of 6.1 mM) during the observed period. In general, a low HGI and a low number of hypoglycaemic events indicate tight and safe blood glucose control. In Cordingley 2009, there were some inconsistencies in the text and tables so conclusions should be made with caution: in one centre (n = 20) the HGI was not significantly different between groups whereas in the other (n = 14) the HGI was significantly greater in the standard care group. In Pachler 2008, the HGI was significantly lower in the computer group (MD ‐1.20 mmol/L, 95% CI ‐1.63 to ‐0.77).
The time in blood glucose range was higher in the computer‐guided glucose management during surgery (Saager 2008: MD 57.00 minutes, 95% CI 9.57 to 104.43), and after surgery in the intensive care unit (Kremen 2007; Saager 2008: pooled MD 257.92 minutes, 95% CI 60.96 to 454.87).
Nine studies reported the mean blood glucose with high heterogeneity (I2 = 86%, one study with opposite direction): the mean blood glucose was significantly lower with the computer advice (pooled SMD ‐0.72, 95% CI ‐1.03 to ‐0.42; pooled MD ‐14.81 mg/dL, 95% CI ‐22.06 to ‐7.56) (Analysis 2.3) (Plank 2006; Augstein 2007; Hovorka 2007; Kremen 2007; Pachler 2008; Saager 2008; Blaha 2009; Wexler 2010; Sato 2011). Excluding two studies with potential unit of analysis error did not change the conclusions (pooled SMD ‐0.55, 95% CI ‐0.83 to ‐0.27; pooled MD ‐10.48 mg/dL, 95% CI ‐17.10 to ‐3.86) (Hovorka 2007; Pachler 2008).
Aminoglycoside antibiotics (five studies)
No data available.
Theophylline (four studies)
No data available.
Anti‐rejection drugs (four studies)
No data available.
Anaesthetic agents (two studies, three comparisons)
No data available.
Antidepressants (one study)
No data available.
Gonadotropins (one study)
No data available.
Summary
In summary, in spite of difference in magnitude effects, computerized advice led more often to a physiological parameter within the desired range for oral anticoagulants (significantly higher percentages of time in target INR range) and for insulin (significantly higher percentages of time in target glucose, lower levels of mean blood glucose).
This result may be explained by the fact that computer advice improved the sampling time and intervals for warfarin or insulin. These two drugs have a narrow therapeutic window. They have variable effects depending on the plasma concentration: a lower dose is ineffective and a higher dose is hazardous. For these drugs, sampling time is critical, since the drug concentration varies over the entire dosing interval and with the duration of dosing in relation to achieving a steady state. Sampling interval, the number of sampling, is also important. For insulin, in three studies, this interval was significantly lower in the computer groups (Hovorka 2007; Pachler 2008; Cordingley 2009), in one study, there was no significant difference between groups (Blaha 2009), whereas in one study, mean sampling intervals were significantly longer in the computer‐assisted group compared with the manual group (Sato 2011).
Hypothesis 3. Decisions on drug dosage based on computer advice led to more rapid therapeutic control, assessed by a physiological parameter
For this comparison, the outcomes analyzed were the time to achieve therapeutic range and the time to stabilization (Analysis 3.1; Analysis 3.2; Figure 11; Figure 12).
Forest plot of comparison: 3 Time to achieve therapeutic control, outcome: 3.1 Time to achieve therapeutic range.
Forest plot of comparison: 3 Time to achieve therapeutic control, outcome: 3.2 Time to stabilization.
Oral anticoagulants (fifteen studies, eighteen comparisons)
Two comparisons analyzed the "time to achieve" therapeutic PT ratio (White 1987; Vadher 1997). There was no evidence of difference between groups (SMD ‐0.22, 95% CI ‐0.69 to 0.26; MD ‐0.58 days, 95% CI ‐1.84 to 0.6).
Three comparisons analyzed the "time to stabilization" (Carter 1987; White 1987; Vadher 1997). In Carter 1987, the outcome was reported for participants who achieved stable PT ratios before discharge. The pooled effect showed a significant reduction in time to achieve stabilization (SMD ‐0.56, 95% CI ‐1.07 to ‐0.04; MD ‐2.49 days, 95% CI ‐3.93 to ‐1.05) even if there was an high inconsistency (I2 = 71%).
Insulin (ten studies)
The "time to target" (time to establish blood glucose control) for insulin was reported in three studies. In one study, the computer advice led to less rapid control of glycaemia (Kremen 2007), whereas there was no significant difference in Blaha 2009 and Cordingley 2009. There was no evidence of difference between groups (SMD 0.22, 95% CI ‐0.52 to 0.95; MD 0.53 hour, 95% CI ‐1.22 to 2.27).
Aminoglycoside antibiotics (five studies)
No data available.
Theophylline (four studies)
No data available.
Anti‐rejection drugs (four studies)
No data available.
Anaesthetic agents (two studies, three comparisons)
No data available.
Antidepressants (one study)
No data available.
Gonadotropins (one study)
No data available.
Summary
The pooled effect showed a significant reduction in time to achieve stabilization for oral anticoagulants (three comparisons) whereas there was no evidence of difference for insulin.
Hypothesis 4. Decisions on drug dosage based on computer advice lead to more effectiveness, assessed by clinical improvement
For this comparison, the outcomes analyzed was the proportion of participants with clinical improvement (Analysis 4.1; Analysis 4.2; Figure 13; Figure 14). Forest plots are only presented for aminoglycoside antibiotics and anti‐rejection drugs (no study or only one for other drugs).
Forest plot of comparison: 4 Clinical improvement, outcome: 4.1 Aminoglycoside antibiotics.
Forest plot of comparison: 4 Clinical improvement, outcome: 4.2 Anti‐rejection drugs.
Anticoagulants (fifteen studies, eighteen comparisons)
No data available.
Insulin (ten studies)
No data available.
Aminoglycoside antibiotics (five studies)
Two studies reported efficacy outcomes. In Burton 1991, there were no statistical differences in the number of participants cured (RR 0.95, 95% CI 0.55 to 1.64), whereas in Leehey 1993, there was more success to respond to treatment in the pharmacist‐direct dosing based on a Bayesian pharmacokinetic dosing program (RR 1.12, 95% CI to 1.03 to 1.23).
Theophylline (four studies)
No data available.
Anti‐rejection drugs (four studies)
Two studies reported efficacy outcomes on biopsy‐confirmed acute rejections. In Asberg 2010, there were no statistical differences in the number of participants without acute rejections (RR 0.94, 95% CI 0.71 to 1.25), whereas in Le Meur 2007, there were more participants with clinical improvement (no acute rejections) in the individualized MMF dosing group (RR 1.22, 95% CI 1.05 to 1.43).
Anaesthetic agents (two studies, three comparisons)
No data available.
Antidepressants (one study)
In Mihajlovic 2003, the participants from individualized or empiric doses of amitriptyline were evaluated clinically using the Hamilton Depression Rating Scale (HAM‐D; 21 items) and Clinical Global Impression Scale (CGI). Total HAM‐D scores were significantly lower in the experimental group after day 28 (MD ‐2.80, 95% CI ‐4.93 to ‐0.67). CGI scores had a statistically significant difference between the groups after day 28 for the three items (in favour of the intervention for Severity of illness and Therapeutic effect, and in favour of the control for Global improvement).
Gonadotropins (one study)
In the study on the effectiveness of a computerized decision support system for ovarian stimulation with gonadotropins (Lesourd 2002), there were no statistical differences in the number of participants with clinical pregnancies (RR 1.15, 95% CI 0.59 to 2.27).
Summary
For each drug, there were, at most, two comparisons. Some comparisons showed significant improvement with computer advice whereas there was no evidence of differences in others.
Hypothesis 5. Decisions on drug dosage based on computer advice lead to fewer unwanted effects
For this comparison, we considered two outcomes: death and adverse reactions.
Death
Ten comparisons analyzed death rates (Analysis 5.1; Figure 15). Globally, there was no significant difference observed between the computer and control groups (RR 1.08, 95% CI 0.80 to 1.45).
Forest plot of comparison: 5 Clinical adverse events, outcome: 5.1 Death.
Clinical adverse events
Nineteen comparisons assessed the proportion of participants with clinical adverse events. Since there was a great diversity of drugs and of type of clinical adverse events, we did not pool the results that were presented by drug.
Anticoagulants (fifteen studies, eighteen comparisons)
Bleeding events
The proportion of participants with bleeding events was available in five comparisons (nine events in computer group, 15 events in control group). In Ageno 2000, the minor bleeding events were observed until discharge or until the seventh day of treatment, in Vadher 1997; Vadher 1997 pop1; Vadher 1997 pop2, haemorrhagic events were collected during the follow‐up (maximum length of follow‐up: 3 to 13 months) and in White 1987, the bleeding complications were collected during hospitalization. There was a trend towards fewer people with bleeding events although it did not reach statistical significance (pooled RR 0.65, 95% CI 0.30 to 1.41) (Analysis 5.2; Figure 16).
Forest plot of comparison: 5 Clinical adverse events, outcome: 5.2 Anticoagulants: events.
The participant exposure time was reported in four comparisons but the bleeding incidence rate could be calculated only in three comparisons (in Fitzmaurice 2000, there was no serious bleeding in the control group). In two studies, there was significantly less bleeding events in the computer group (Claes 2005; Poller 2008 PARMA 5), whereas the difference was not statistically significant in one study (Poller 2009 DAWN AC). The pooled effect showed a non‐significant reduction in bleeding events with an estimated rate ratio of 0.81 (95% CI 0.60 to 1.08) (Analysis 5.3; Figure 17).
Forest plot of comparison: 5 Clinical adverse events, outcome: 5.3 Anticoagulants: event rates.
For heparin, there was no statistical difference in bleeding events between the computer‐assisted and the nomogram‐directed therapy (Mungall 1994).
Thromboembolism
In two studies, there was no thromboembolism due to undertreatment in both computer and standard dosing groups (Vadher 1997 pop2; Mitra 2005). In two studies, the number of participants with thromboembolism tended to be higher in the computer group but the pooled effect showed no statistical difference between groups (pooled RR 3.25, 95% CI 0.66 to 16.03) (Vadher 1997; Vadher 1997 pop1) (Analysis 5.2; Figure 16).
For the incidence rate, in two studies, there was significantly less thromboembolism events in the computer group (Fitzmaurice 2000; Claes 2005), whereas the difference was not significant in two studies (Poller 2008 PARMA 5; Poller 2009 DAWN AC). The pooled effect showed a significant reduction in thromboembolism events with an estimated rate ratio of 0.68 (95% CI 0.49 to 0.94) but the inconsistency was high (I2 = 65%) (Analysis 5.3; Figure 17).
Total clinical adverse events
In Anderson 2007, total clinical adverse events (clinical events plus INR 4 or greater) were numerically fewer in the pharmacogenetic than standard arm (34 versus 42), although the difference was not significant (P value = 0.26) and the serious clinical events were infrequent (pharmacogenetic 4, standard 5) and were unrelated to out‐of‐range INRs.
In Mungall 1994 (heparin), there was a trend towards less clinical adverse events in the computer‐assisted heparin therapy compared with nomogram‐directed therapy (RR 0.08, 95% CI 0.00 to 1.35).
Insulin (ten studies)
Hypoglycaemia
The proportion of participants with hypoglycaemia was available in seven comparisons (11 events in computer group, 17 events in control group). Hypoglycaemia was defined as blood glucose less than 60 mg/dL (< 3.3 mmol/L) in three studies (Saager 2008; Cordingley 2009; Wexler 2010), less than less 54 mg/dL (< 3.0 mmol/L) in one study (Plank 2006), and less than 52 mg/dL (< 2.9 mmol/L) in three studies (Hovorka 2007; Kremen 2007; Sato 2011). No significant difference was observed between the computer and control groups (pooled RR 0.71, 95% CI 0.35 to 1.48) (Analysis 5.4; Figure 18).
Forest plot of comparison: 5 Clinical adverse events, outcome: 5.4 Insulin.
Severe hypoglycaemia
The proportion of participants with severe hypoglycaemia was available in four comparisons (one event in computer group, two events in control group). Severe hypoglycaemia was defined as blood glucose less than 40 mg/dL (< 2.2 mmol/L) in three studies (Pachler 2008; Cordingley 2009; Wexler 2010), and less than less 41 mg/dL (< 2.3 mmol/L) in one study (Blaha 2009), and less than 42 mg/dL (<2.9 mmol/L) in three studies (Hovorka 2007; Kremen 2007; Sato 2011). No significant difference was observed between the computer and control groups (pooled RR 0.69, 95% CI 0.11 to 4.31).
Aminoglycoside antibiotics (five studies)
Nephrotoxicity
Four studies reported outcomes on nephrotoxicity (Begg 1989; Destache 1990; Burton 1991; Leehey 1993). There was a trend towards lower nephrotoxicity in computer group although the difference was not significant (pooled RR 0.67, 95% CI 0.42 to 1.06) (Analysis 5.5; Figure 19).
Forest plot of comparison: 5 Clinical adverse events, outcome: 5.5 Aminoglycoside antibiotics.
Need for dialysis
Leehey 1993 compared people who needed dialysis. No difference was observed between the pharmacist‐directed dosing using a Bayesian pharmacokinetic dosing program and the control group (RR 0.30, 95% CI 0.03 to 2.86).
Theophylline (four studies)
Tachycardia
Casner 1993 analyzed the theophylline toxicity (including nausea, vomiting, tremor, tachycardia and seizures): one participant had a tachycardia in the kinetic group whereas there was no episode of toxicity in the empiric group (RR 3.17, 95% CI 0.14 to 72.80).
Total clinical adverse events
In Gonzalez 1989, adverse reactions occurred in 7% of participants with population‐based emergency department guidelines dosing, and 10% of participants with Bayesian‐derived pharmacokinetic dosing (RR 1.62, 95% CI 0.32 to 8.26).
Anti‐rejection drugs (four studies)
Cytomegalovirus infections
Two comparisons reported outcomes on cytomegalovirus infections (Le Meur 2007; Asberg 2010). No significant difference was observed between the computer and control groups (pooled RR 0.90, 95% CI 0.58 to 1.40) (Analysis 5.6; Figure 20).
Forest plot of comparison: 5 Clinical adverse events, outcome: 5.6 Anti‐rejection drugs.
Total clinical adverse events
Le Meur 2007 analyzed clinical adverse events including anaemia, leukopenia, gastrointestinal adverse events and infections. There was no statistical difference between individualized doses based on therapeutic monitoring of MPA and fixed‐dose MMF (RR 1.07, 95% CI 0.98 to 1.17).
Anaesthetic agents (two studies, three comparisons)
Total clinical adverse events
In Rodman 1984, there was no clinical adverse event in the people monitored clinically (monitoring of rhythm, intermittent hard‐copy rhythm strips, serial ECGs, daily measurements of electrolyte and cardiac enzyme levels, liver function tests).
Antidepressants (one study)
Total clinical adverse events
The comparison of safety between individualized and empiric dose regimen of amitriptyline in the treatment of major depressive episode (Mihajlovic 2003) was studied in Mihajlovic 2010. The CGI scale and originally designed questionnaire were used for clinical adverse events assessment. In the experimental group, 69 complaints on nine different types of adverse effects were recorded during the eight‐week treatment period and in control group, 111 complaints on 12 different types of adverse effects were recorded. Mihajlovic 2010 indicated that "significantly higher number of patients complaining on adverse effects were in the control group". Nevertheless, none of the comparisons during the eight‐week treatment period was significant (RR: day 14 0.63, 95% CI 0.34 to 1.15; day 28 0.71, 95% CI 0.41 to 1.21; day 42 0.75, 95% CI 0.30 to 1.90; day 56 0.75, 95% CI 0.30 to 1.90).
Gonadotropins (one study)
No data available.
Summary
No significant difference on death was observed between the computer and control groups. For clinical adverse events, we did not pool the results because of the diversity of outcomes. There was a trend for less nephrotoxicity for aminoglycoside antibiotics, for less bleeding and thromboembolism events for anticoagulants with the computer group. When considering the incidence rates for anticoagulants, which is a more precise outcome measure taking into account the exposure time, this difference was significant in favour of the computer group for thromboembolism despite a high heterogeneity. There was no evidence of difference for insulin, theophylline, anaesthetic agents, anti‐rejection drugs and antidepressants.
Hypothesis 6. Computer advice reduces the cost of health care or the use of resources (length of stay)
For this comparison, the outcomes analyzed were the length of stay (Analysis 6.1; Figure 21) and the cost per participant.
Forest plot of comparison: 6 Healthcare resources, outcome: 6.1 Length of stay (days).
Length of stay
Nine comparisons analyzed the length of stay (oral anticoagulants: White 1987; Mitra 2005; insulin: Wexler 2010; aminoglycoside: Destache 1990; Burton 1991; theophylline: Hurley 1986; Verner 1992; Casner 1993; cyclosporine: Chertow 2001). There was a significant reduction of the length of stay in the computer group in four comparisons (Hurley 1986; White 1987; Burton 1991; Chertow 2001), whereas the difference was not significant in five comparisons. The pooled effect tended to be in favour of the computer group (SMD ‐0.15, 95% CI ‐0.33 to 0.02).
Costs per participant and incremental cost‐effectiveness ratio
Five studies analyzed the costs of interventions using computer advice on drug dosage.
In Destache 1990, clinical pharmacokinetic service direct costs were significantly lower than usual care (without clinical pharmacokinetic service monitoring) for aminoglycoside antibiotics: USD7102.56 ± 8898.18 compared with USD13,758.64 ± 22,874.31 (P value < 0.02).
In Chertow 2001, there were no significant differences between intervention and control periods in estimated hospital and pharmacy costs for cyclosporine.
In Claes 2006 (main results in Claes 2005), the total cost per participant per month was EUR53.79 for usual care, EUR50.62 for intervention A (multifaceted intervention) and EUR53.20 for intervention D (multifaceted intervention plus DAWN AC computer advice) for anticoagulants.
Jowett 2009 reported the cost‐effectiveness for anticoagulants of the randomized multicentre study of two computer‐assisted dosage programs (DAWN AC or PARMA 5) versus manual dosing conducted by Poller et al. (Poller 2008 PARMA 5; Poller 2009 DAWN AC). Dosing time and costs were available in 28 of the 32 clinics participating. Total overall costs per participant were significantly lower in the computer‐dosing arm (EUR ‐50.5, bootstrapped 95% CI ‐76.8 to ‐24.1), mainly driven by the difference in dosing costs. The authors concluded that computer‐assisted dosage with the two programs (DAWN AC and PARMA 5) was not less effective clinically but was lower in cost than manual dosage.
In Rousseau 2010 (main results in Le Meur 2007), the mean total yearly cost per participant was EUR47,477 (95% CI 43,933 to 51,020) in the concentration‐controlled group and EUR46,783 (95% CI 44,152 to 49,414) in the fixed‐dose group (P value = 0.7) for mycophenolate.
In three studies (Claes 2006; Jowett 2009; Rousseau 2010), the incremental cost‐effectiveness ratio (ICER) was reported. The ICER represents the additional cost required to provide one unit of additional effect. In Jowett 2009, as the costs were lower and the intervention more favourable, calculation of an ICER in this case was not appropriate. In Rousseau 2010, the incremental 12‐month cost was EUR3757 per treatment failure (Purchasing Power Parities United States/France: USD4129). In Claes 2006, the ICER for intervention A (multifaceted intervention) was EUR5.2 per day within range (DWR) and for intervention D (multifaceted intervention plus DAWN AC computer advice) EUR4.9 per DWR. Intervention A was less effective but also slightly less costly compared with D resulting in comparable cost‐effectiveness ratios.
Summary
There was a trend to a reduction of the length of stay in the computer groups. In most of trials, the computer‐assisted dosage was not less effective but was lower in cost than manual dosage resulting in a comparable or better cost‐effectiveness ratio than usual care.
Subgroup analyses
No study compared the effect of decision support logistics or organizations of care. We found no studies where the computerized advice was delivered by delayed feedback.
Since fewer than 10 studies were available for each characteristic to model, we did not investigate heterogeneity.
Discussion
Summary of main results
This update found similar results to the previous review (Durieux 2008), and, in addition, specific therapeutic areas where the intervention is beneficial (the level of evidence was based upon the 'Summary of findings' tables):
1. For oral anticoagulants, CDSS led more often to an INR within the desired range and reduced the time to achieve stabilization of PT and the incidence rates of thromboembolism, in a statistical significant manner.
2. For insulin, it significantly improved the percentage of time in target glucose and reduced the mean blood glucose. For clinical outcomes, there was no efficacy outcome assessed and the difference for safety outcomes was not statistically significant.
3. For aminoglycoside antibiotics, it was significantly more efficient for reaching appropriate drug serum concentrations. For clinical outcomes, positive treatment response was inconsistent in the two studies.
For all these drugs, the quality of evidence was very low due to the high heterogeneity, the low quality and small sample size in most of studies.
4. For the other drugs (theophylline, anti‐rejection drugs, anaesthetic agents, antidepressants and gonadotropins), the level of evidence was too low (small number and sample size of studies) to conclude.
5. Overall, CDSS tended to reduce the length of stay but the difference did not reach statistical significance. Quality of evidence by drugs was low or moderate since there were generally only one or two trials studying this outcome. In most trials, the computer‐assisted dosage had lower or equivalent costs than manual dosage.
Overall completeness and applicability of evidence
This review included 46 comparisons, of which 18 (39%) concerned anticoagulants and 10 (22%) insulin dosage, providing more consistent results for these drugs than for others (fewer than five comparisons). Five comparisons (10%) concerned aminoglycoside antibiotics but all studies were published before 1993. Three comparisons concerned theophylline, a drug that is not considered as the first‐choice treatment of asthma at present. However, monitoring serum concentrations of theophylline is essential to ensure that non‐toxic doses are achieved (National Asthma 2002). Compared with the last version of this review, our update included new drugs such as insulin and anti‐rejection drugs. We found no studies concerning some new drugs for which it is considered important to monitor drug levels such as glycopeptides, antifungal (fluconazole) and antiretroviral drugs. CDSS may be more efficient in sampling time and intervals for drugs with a narrow therapeutic window such as insulin or oral anticoagulants. However, we found no studies for other drugs with narrow therapeutic indexes (such as lithium, vancomycin, levothyroxine, digoxin, carbamazepine and phenytoin).
Quality of the evidence
However, the findings need to be read with caution.
First, the quality of studies was generally low (see 'Summary of findings' tables). There was no blinding of participants and personnel in all studies. The random sequence generation or the allocation concealment, or both were unclear in half of the studies. In most comparisons, sample size was small.
Second, even when grouping the studies by drug, heterogeneity remained high because of the widely different outcome definitions, clinical contexts or organizations of care. This issue limits the comparability between studies and the interpretation of the clinical significance of the results presented with the SMDs. Besides, the intervention type varied between studies. For example, Anderson 2007 used a pharmacogenetic guidance program for warfarin initiation. This differs totally with all the other studies on oral anticoagulants.
Third, for some indicators (length of stay, mortality), crude results can be affected by unknown confounding factors.
Last, the results are significant for relevant physiological outcomes (assessed for oral anticoagulants and insulin). However, efficacy clinical outcomes were assessed only for four drugs (aminoglycoside antibiotics, anti‐rejection drugs, antidepressants and gonadotropins) and there were at most two comparisons for each drug. For aminoglycoside antibiotics and anti‐rejection drugs (two comparisons), the studies showed contradictory results.
For safety clinical events, no significant difference on death was observed between the computer and control groups. For clinical adverse events, there was a trend for less nephrotoxicity for aminoglycoside antibiotics, for fewer bleeding and thromboembolism events for anticoagulants with the computer group. When considering the incidence rates for thromboembolism, this difference was significantly in favour of the computer group for anticoagulants despite a high heterogeneity. There was no significant difference in clinical adverse events for the other drugs.
Potential biases in the review process
We identified no potential biases. We considered congress reports and other sources of unpublished studies.
Agreements and disagreements with other studies or reviews
Nine reviews assessed the effect of computer advice on drug dosage. All found a low methodological quality and a high heterogeneity of studies included. Most presented a publication bias or no evaluation of the methodological quality of the studies included. We could not compare our results with seven reviews that reported their quantitative results as the percentage of studies that showed a significant improvement with computer advice (Garg 2005 updated by Nieuwlaat 2011a; Yourman 2008; Eslami 2009; Mollon 2009; Pearson 2009; Robertson 2010).
Two old reviews (Chatellier 1998; Fitzmaurice 1998a) suggested that computer advice may improve therapeutic INR control for oral anticoagulants, but no meta‐analysis was performed and the evaluation of the quality of studies included was not clear or adequate.
Eslami 2009 performed a systematic review on tight glycaemic control with insulin in intensive care units, but performed no meta‐analysis because of the high heterogeneity. The results were expressed as the percentage of studies that showed a significant improvement with computer advice on the blood glucose regulatory process. Although most studies reported a positive effect, the evidence was very low: the review only searched for published studies in MEDLINE, there was no evaluation of the methodological quality of studies, which were mostly before‐after designs. Causality was difficult to determine because of the simultaneous implementation of a new tight glycaemic control protocol with computer advice.
Three other reviews assessed the effect of computer advice more generally on all drugs, focusing on drug dosage, but their results were expressed as the percentage of studies that showed a significant improvement with computer advice (Mollon 2009; Pearson 2009; Robertson 2010). Mollon 2009 excluded all RCTs that focused only on dose adjustment.
Two reviews assessed the general benefits of computer advice: Yourman 2008 for improving medication prescribing in older adults and Garg 2005, updated by Nieuwlaat 2011a, for drug monitoring and dosing. Unlike us, their results were the percentage of studies that showed a significant improvement with computer advice. A study was considered to have a positive effect (i.e. CDSS showed improvement) if at least 50% of the relevant study outcomes were statistically significantly positive. This approach does not give insight in the magnitude of effects and may have underestimated the overall efficacy. In contrast, there is a risk for publication bias of positive RCTs, which could cause overestimation of CDSS efficacy. Yourman 2008 found that computer advice was effective in improving medication prescribing in older adults, and Nieuwlaat 2011a in improving the process of care. In Yourman 2008, there was no evaluation of the methodological quality of included studies, a likely publication bias and a high heterogeneity. In Nieuwlaat 2011a, the studies were small and of low quality when using a 10‐point‐scale extended from the Jadad scale.
Study flow diagram.
Risk of bias graph: review authors' judgements about each risk of bias item presented as percentages across all included studies.
Risk of bias summary: review authors' judgements about each risk of bias item for each included study.
Forest plot of comparison: 1 Serum concentrations and therapeutic range, outcome: 1.1 Serum concentrations (mg/L) ‐ part A (SMD > 0 in favour of the intervention).
Forest plot of comparison: 1 Serum concentrations and therapeutic range, outcome: 1.2 Serum concentrations (ng/L) ‐ part B (SMD < 0 in favour of the intervention).
Forest plot of comparison: 1 Serum concentrations and therapeutic range, outcome: 1.3 Proportion of participants within therapeutic range.
Forest plot of comparison: 1 Serum concentrations and therapeutic range, outcome: 1.4 Proportion of participants with toxic drug levels.
Forest plot of comparison: 2 Physiological parameters, outcome: 2.1 Oral anticoagulants: % time in target INR range.
Forest plot of comparison: 2 Physiological parameters, outcome: 2.2 Insulin: % time in target glucose range.
Forest plot of comparison: 2 Physiological parameters, outcome: 2.3 Insulin: mean blood glucose (mg/dL).
Forest plot of comparison: 3 Time to achieve therapeutic control, outcome: 3.1 Time to achieve therapeutic range.
Forest plot of comparison: 3 Time to achieve therapeutic control, outcome: 3.2 Time to stabilization.
Forest plot of comparison: 4 Clinical improvement, outcome: 4.1 Aminoglycoside antibiotics.
Forest plot of comparison: 4 Clinical improvement, outcome: 4.2 Anti‐rejection drugs.
Forest plot of comparison: 5 Clinical adverse events, outcome: 5.1 Death.
Forest plot of comparison: 5 Clinical adverse events, outcome: 5.2 Anticoagulants: events.
Forest plot of comparison: 5 Clinical adverse events, outcome: 5.3 Anticoagulants: event rates.
Forest plot of comparison: 5 Clinical adverse events, outcome: 5.4 Insulin.
Forest plot of comparison: 5 Clinical adverse events, outcome: 5.5 Aminoglycoside antibiotics.
Forest plot of comparison: 5 Clinical adverse events, outcome: 5.6 Anti‐rejection drugs.
Forest plot of comparison: 6 Healthcare resources, outcome: 6.1 Length of stay (days).
Comparison 1 Serum concentrations and therapeutic range, Outcome 1 Serum concentrations (mg/L) ‐ part A (SMD > 0 in favour of the intervention).
Comparison 1 Serum concentrations and therapeutic range, Outcome 2 Serum concentrations (ng/L) ‐ part B (SMD < 0 in favour of the intervention).
Comparison 1 Serum concentrations and therapeutic range, Outcome 3 Proportion of participants within therapeutic range.
Comparison 1 Serum concentrations and therapeutic range, Outcome 4 Proportion of participants with toxic drug levels.
Comparison 2 Physiological parameters, Outcome 1 Oral anticoagulants: % time in target INR range.
Comparison 2 Physiological parameters, Outcome 2 Insulin: % time in target glucose range.
Comparison 2 Physiological parameters, Outcome 3 Insulin: mean blood glucose (mg/dL).
Comparison 3 Time to achieve therapeutic control, Outcome 1 Time to achieve therapeutic range.
Comparison 3 Time to achieve therapeutic control, Outcome 2 Time to stabilization.
Comparison 4 Clinical improvement, Outcome 1 Aminoglycoside antibiotics.
Comparison 4 Clinical improvement, Outcome 2 Anti‐rejection drugs.
Comparison 5 Clinical adverse events, Outcome 1 Death.
Comparison 5 Clinical adverse events, Outcome 2 Anticoagulants: events.
Comparison 5 Clinical adverse events, Outcome 3 Anticoagulants: event rates.
Comparison 5 Clinical adverse events, Outcome 4 Insulin.
Comparison 5 Clinical adverse events, Outcome 5 Aminoglycoside antibiotics.
Comparison 5 Clinical adverse events, Outcome 6 Anti‐rejection drugs.
Comparison 6 Healthcare resources, Outcome 1 Length of stay (days).
| Computerized advice on drug dosage for leading serum concentrations within therapeutic range | ||||||
| Patient or population: patients with leading serum concentrations within therapeutic range | ||||||
| Outcomes | Illustrative comparative risks* (95% CI) | Relative effect | No of participants | Quality of the evidence | Comments | |
| Assumed risk | Corresponding risk | |||||
| Control | Computerized advice on drug dosage | |||||
| Serum concentrations ‐ aminoglycoside antibiotics: peak concentration | ‐ | The mean serum concentrations ‐ aminoglycoside antibiotics: peak concentration in the intervention groups was | ‐ | 372 | ⊕⊕⊝⊝ | SMD 0.79 (95% CI 0.46 to 1.13) |
| Serum concentrations ‐ theophylline | ‐ | The mean serum concentrations ‐ theophylline in the intervention groups was | ‐ | 201 | ⊕⊕⊝⊝ | SMD 0.41 (95% CI ‐0.2 to 1.02) |
| Proportion of participants within therapeutic range ‐ aminoglycoside antibiotics: % of participants with peak concentrations adequate after 2 days | Study population | RR 4.44 | 72 | ⊕⊕⊕⊝ | ‐ | |
| 135 per 1000 | 600 per 1000 | |||||
| Moderate | ||||||
| 151 per 1000 | 670 per 1000 | |||||
| Proportion of participants with toxic drug levels ‐ theophylline | Study population | RR 0.53 | 109 | ⊕⊕⊕⊝ | ‐ | |
| 273 per 1000 | 145 per 1000 | |||||
| Moderate | ||||||
| 217 per 1000 | 115 per 1000 | |||||
| *The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). | ||||||
| GRADE Working Group grades of evidence | ||||||
| 1 Lack of blinding of participants and personnel in all studies. Incomplete outcome data in three studies. Random sequence generation and allocation concealment unclear in one study. | ||||||
| Computerized advice on drug dosage for leading physiological parameters within therapeutic range | ||||||
| Patient or population: patients with leading physiological parameters within therapeutic range | ||||||
| Outcomes | Illustrative comparative risks* (95% CI) | Relative effect | No of participants | Quality of the evidence | Comments | |
| Assumed risk | Corresponding risk | |||||
| Control | Computerized advice on drug dosage | |||||
| Oral anticoagulants: time in target INR range (%) | ‐ | The mean oral anticoagulants: time in target INR range (%) in the intervention groups was | ‐ | 13,581 | ⊕⊝⊝⊝ | SMD 0.19 (95% CI 0.06 to 0.33) |
| Insulin: time in target glucose range (%) | ‐ | The mean insulin: time in target glucose range (%) in the intervention groups was | ‐ | 234 | ⊕⊕⊝⊝ | SMD 1.27 (95% CI 0.56 to 1.98) |
| *The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). | ||||||
| GRADE Working Group grades of evidence | ||||||
| 1 No information given on random sequence generation and allocation concealment in half of the studies. | ||||||
| Computerized advice on drug dosage for reducing time to achieve therapeutic control | ||||||
| Patient or population: patients with reducing time to achieve therapeutic control | ||||||
| Outcomes | Illustrative comparative risks* (95% CI) | Relative effect | No of participants | Quality of the evidence | Comments | |
| Assumed risk | Corresponding risk | |||||
| Control | Computerized advice on drug dosage | |||||
| Time to achieve therapeutic range ‐ oral anticoagulants: time to achieve therapeutic prothrombin ratio (days) | ‐ | The mean time to achieve therapeutic range ‐ oral anticoagulants: time to achieve therapeutic prothrombin ratio (days) in the intervention groups was | ‐ | 223 | ⊕⊕⊝⊝ | SMD ‐0.22 (95% CI ‐0.69 to 0.26) |
| Time to achieve therapeutic range ‐ insulin: time to achieve therapeutic control (hours) | ‐ | The mean time to achieve therapeutic range ‐ insulin: time to achieve therapeutic control (hours) in the intervention groups was | ‐ | 194 | ⊕⊝⊝⊝ | SMD ‐0.14 (95% CI ‐0.98 to 0.7) |
| Time to stabilization ‐ oral anticoagulants: time to stabilization (days) | ‐ | The mean time to stabilization ‐ oral anticoagulants: time to stabilization (days) in the intervention groups was | ‐ | 255 | ⊕⊝⊝⊝ | SMD ‐0.56 (95% CI ‐1.07 to ‐0.04) |
| *The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). | ||||||
| GRADE Working Group grades of evidence | ||||||
| 1 Allocation concealment was unclear, and there was no blinding of participants and personnel in both studies. | ||||||
| Computerized advice on drug dosage for leading to fewer clinical adverse events | ||||||
| Patient or population: patients with leading to fewer clinical adverse events | ||||||
| Outcomes | Illustrative comparative risks* (95% CI) | Relative effect | No of participants | Quality of the evidence | Comments | |
| Assumed risk | Corresponding risk | |||||
| Control | Computerized advice on drug dosage | |||||
| Death | Study population | RR 1.08 | 14,046 | ⊕⊕⊝⊝ | ‐ | |
| 12 per 1000 | 13 per 1000 | |||||
| Moderate | ||||||
| 28 per 1000 | 30 per 1000 | |||||
| Anticoagulants: events ‐ bleeding | Study population | RR 0.65 | 552 | ⊕⊕⊝⊝ | ‐ | |
| 61 per 1000 | 40 per 1000 | |||||
| Moderate | ||||||
| 65 per 1000 | 42 per 1000 | |||||
| Anticoagulants: events ‐ thromboembolism | Study population | RR 3.25 | 355 | ⊕⊝⊝⊝ | ‐ | |
| 11 per 1000 | 36 per 1000 | |||||
| Moderate | ||||||
| 7 per 1000 | 23 per 1000 | |||||
| Insulin ‐ hypoglycaemia (< 60 mg/dL) | Study population | RR 0.71 | 378 | ⊕⊕⊝⊝ | ‐ | |
| 90 per 1000 | 64 per 1000 | |||||
| Moderate | ||||||
| 67 per 1000 | 48 per 1000 | |||||
| Insulin ‐ severe hypoglycaemia (< 40 mg/dL) | Study population | RR 0.69 | 292 | ⊕⊝⊝⊝ | ‐ | |
| 14 per 1000 | 9 per 1000 | |||||
| Moderate | ||||||
| 8 per 1000 | 6 per 1000 | |||||
| Aminoglycoside antibiotics ‐ nephrotoxicity | Study population | RR 0.67 | 493 | ⊕⊕⊝⊝ | ‐ | |
| 162 per 1000 | 108 per 1000 | |||||
| Moderate | ||||||
| 154 per 1000 | 103 per 1000 | |||||
| *The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). | ||||||
| GRADE Working Group grades of evidence | ||||||
| 1 No (or unclear) blinding of participants and personnel in all studies. In half of the studies: sequence generation and/or allocation concealment were unclear, and the data were incomplete. | ||||||
| Saving healthcare resources for saving healthcare resources | ||||||
| Patient or population: saving healthcare resources | ||||||
| Outcomes | Illustrative comparative risks* (95% CI) | Relative effect | No of participants | Quality of the evidence | Comments | |
| Assumed risk | Corresponding risk | |||||
| Control | Saving healthcare resources | |||||
| Length of stay (days) | ‐ | The mean length of stay (days) in the intervention groups was | ‐ | 18,507 | ⊕⊝⊝⊝ | SMD ‐0.15 (95% CI ‐0.33 to 0.02) |
| Length of stay (days) ‐ oral anticoagulants | ‐ | The mean length of stay (days) ‐ oral anticoagulants in the intervention groups was | ‐ | 105 | ⊕⊝⊝⊝ | SMD ‐0.12 (95% CI ‐1.1 to 0.86) |
| Length of stay (days) ‐ insulin | ‐ | The mean length of stay (days) ‐ insulin in the intervention groups was | ‐ | 128 | ⊕⊕⊕⊕ | SMD 0.18 (95% CI ‐0.17 to 0.53) |
| Length of stay (days) ‐ theophylline | ‐ | The mean length of stay (days) ‐ theophylline in the intervention groups was | ‐ | 151 | ⊕⊕⊝⊝ | SMD ‐0.2 (95% CI ‐0.56 to 0.16) |
| Length of stay (days) ‐ aminoglycoside antibiotics | ‐ | The mean length of stay (days) ‐ aminoglycoside antibiotics in the intervention groups was | ‐ | 295 | ⊕⊕⊕⊝ | SMD ‐0.35 (95% CI ‐0.58 to ‐0.12) |
| Length of stay (days) ‐ anti‐rejection drugs | ‐ | The mean length of stay (days) ‐ anti‐rejection drugs in the intervention groups was | ‐ | 17,828 | ⊕⊕⊕⊝ | SMD ‐0.04 (95% CI ‐0.07 to ‐0.01) |
| *The basis for the assumed risk (e.g. the median control group risk across studies) is provided in footnotes. The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). | ||||||
| GRADE Working Group grades of evidence | ||||||
| 1 I2 = 57%. | ||||||
| Outcome or subgroup title | No. of studies | No. of participants | Statistical method | Effect size |
| 1 Serum concentrations (mg/L) ‐ part A (SMD > 0 in favour of the intervention) Show forest plot | 9 | Std. Mean Difference (IV, Random, 95% CI) | Subtotals only | |
| 1.1 Aminoglycoside antibiotics: peak concentration | 4 | 372 | Std. Mean Difference (IV, Random, 95% CI) | 0.79 [0.46, 1.13] |
| 1.2 Theophylline | 4 | 201 | Std. Mean Difference (IV, Random, 95% CI) | 0.41 [‐0.20, 1.02] |
| 1.3 Lidocaine | 1 | 20 | Std. Mean Difference (IV, Random, 95% CI) | 1.32 [0.33, 2.32] |
| 2 Serum concentrations (ng/L) ‐ part B (SMD < 0 in favour of the intervention) Show forest plot | 3 | Std. Mean Difference (IV, Random, 95% CI) | Subtotals only | |
| 2.1 Midazolam | 1 | 24 | Std. Mean Difference (IV, Random, 95% CI) | ‐1.43 [‐2.34, ‐0.51] |
| 2.2 Fentanyl | 1 | 24 | Std. Mean Difference (IV, Random, 95% CI) | 0.27 [‐0.53, 1.08] |
| 2.3 Antidepressants: steady‐state plasma concentration | 1 | 60 | Std. Mean Difference (IV, Random, 95% CI) | ‐0.68 [‐1.20, ‐0.16] |
| 3 Proportion of participants within therapeutic range Show forest plot | 3 | Risk Ratio (M‐H, Random, 95% CI) | Subtotals only | |
| 3.1 Aminoglycoside antibiotics: % of participants with peak concentrations adequate after 2 days | 2 | 72 | Risk Ratio (M‐H, Random, 95% CI) | 4.44 [1.94, 10.13] |
| 3.2 Anti‐rejection drugs | 1 | 125 | Risk Ratio (M‐H, Random, 95% CI) | 0.71 [0.38, 1.32] |
| 4 Proportion of participants with toxic drug levels Show forest plot | 2 | Risk Ratio (M‐H, Random, 95% CI) | Subtotals only | |
| 4.1 Theophylline | 2 | 109 | Risk Ratio (M‐H, Random, 95% CI) | 0.53 [0.25, 1.13] |
| Outcome or subgroup title | No. of studies | No. of participants | Statistical method | Effect size |
| 1 Oral anticoagulants: % time in target INR range Show forest plot | 6 | 13581 | Mean Difference (IV, Random, 95% CI) | 3.68 [0.90, 6.45] |
| 2 Insulin: % time in target glucose range Show forest plot | 4 | 234 | Mean Difference (IV, Random, 95% CI) | 22.18 [9.94, 34.43] |
| 3 Insulin: mean blood glucose (mg/dL) Show forest plot | 9 | 520 | Std. Mean Difference (IV, Random, 95% CI) | ‐0.72 [‐1.03, ‐0.42] |
| Outcome or subgroup title | No. of studies | No. of participants | Statistical method | Effect size |
| 1 Time to achieve therapeutic range Show forest plot | 5 | Mean Difference (IV, Random, 95% CI) | Subtotals only | |
| 1.1 Oral anticoagulants: time to achieve therapeutic prothrombin ratio (days) | 2 | 223 | Mean Difference (IV, Random, 95% CI) | ‐0.58 [‐1.84, 0.69] |
| 1.2 Insulin: time to achieve therapeutic control (h) | 3 | 134 | Mean Difference (IV, Random, 95% CI) | 0.53 [‐1.22, 2.27] |
| 2 Time to stabilization Show forest plot | 3 | Std. Mean Difference (IV, Random, 95% CI) | Subtotals only | |
| 2.1 Oral anticoagulants: time to stabilization (days) | 3 | 255 | Std. Mean Difference (IV, Random, 95% CI) | ‐0.56 [‐1.07, ‐0.04] |
| Outcome or subgroup title | No. of studies | No. of participants | Statistical method | Effect size |
| 1 Aminoglycoside antibiotics Show forest plot | 2 | Risk Ratio (M‐H, Random, 95% CI) | Subtotals only | |
| 1.1 Number of participants cured | 2 | 271 | Risk Ratio (M‐H, Random, 95% CI) | 1.12 [1.02, 1.22] |
| 2 Anti‐rejection drugs Show forest plot | 2 | Risk Ratio (M‐H, Fixed, 95% CI) | Subtotals only | |
| 2.1 No biopsy‐confirmed rejections | 2 | 170 | Risk Ratio (M‐H, Fixed, 95% CI) | 1.15 [1.00, 1.32] |
| Outcome or subgroup title | No. of studies | No. of participants | Statistical method | Effect size |
| 1 Death Show forest plot | 10 | 14046 | Risk Ratio (M‐H, Random, 95% CI) | 1.08 [0.80, 1.45] |
| 1.1 Oral anticoagulants | 5 | 13499 | Risk Ratio (M‐H, Random, 95% CI) | 1.09 [0.78, 1.51] |
| 1.2 Aminoglycoside antibiotics | 3 | 326 | Risk Ratio (M‐H, Random, 95% CI) | 0.66 [0.14, 3.10] |
| 1.3 Theophylline | 1 | 91 | Risk Ratio (M‐H, Random, 95% CI) | 0.18 [0.01, 3.64] |
| 1.4 Cyclosporine | 1 | 130 | Risk Ratio (M‐H, Random, 95% CI) | 1.0 [0.06, 15.65] |
| 2 Anticoagulants: events Show forest plot | 7 | Risk Ratio (M‐H, Random, 95% CI) | Subtotals only | |
| 2.1 Bleeding | 6 | 552 | Risk Ratio (M‐H, Random, 95% CI) | 0.65 [0.30, 1.41] |
| 2.2 Thromboembolism | 4 | 355 | Risk Ratio (M‐H, Random, 95% CI) | 3.25 [0.66, 16.03] |
| 3 Anticoagulants: event rates Show forest plot | 4 | Rate Ratio (Random, 95% CI) | Subtotals only | |
| 3.1 Bleeding | 4 | 18902 | Rate Ratio (Random, 95% CI) | 0.81 [0.60, 1.08] |
| 3.2 Thromboembolism | 4 | 18902 | Rate Ratio (Random, 95% CI) | 0.68 [0.49, 0.94] |
| 4 Insulin Show forest plot | 9 | Risk Ratio (M‐H, Random, 95% CI) | Subtotals only | |
| 4.1 Hypoglycaemia (< 60 mg/dL) | 7 | 378 | Risk Ratio (M‐H, Random, 95% CI) | 0.71 [0.35, 1.48] |
| 4.2 Severe hypoglycaemia (< 40 mg/dL) | 4 | 292 | Risk Ratio (M‐H, Random, 95% CI) | 0.69 [0.11, 4.31] |
| 5 Aminoglycoside antibiotics Show forest plot | 4 | Risk Ratio (M‐H, Random, 95% CI) | Subtotals only | |
| 5.1 Nephrotoxicity | 4 | 493 | Risk Ratio (M‐H, Random, 95% CI) | 0.67 [0.42, 1.06] |
| 6 Anti‐rejection drugs Show forest plot | 2 | Risk Ratio (M‐H, Random, 95% CI) | Subtotals only | |
| 6.1 Cytomegalovirus infections | 2 | 170 | Risk Ratio (M‐H, Random, 95% CI) | 0.90 [0.58, 1.40] |
| Outcome or subgroup title | No. of studies | No. of participants | Statistical method | Effect size |
| 1 Length of stay (days) Show forest plot | 9 | 18507 | Std. Mean Difference (IV, Random, 95% CI) | ‐0.15 [‐0.33, 0.02] |
| 1.1 Oral anticoagulants | 2 | 105 | Std. Mean Difference (IV, Random, 95% CI) | ‐0.12 [‐1.10, 0.86] |
| 1.2 Insulin | 1 | 128 | Std. Mean Difference (IV, Random, 95% CI) | 0.18 [‐0.17, 0.53] |
| 1.3 Theophylline | 3 | 151 | Std. Mean Difference (IV, Random, 95% CI) | ‐0.20 [‐0.56, 0.16] |
| 1.4 Aminoglycoside antibiotics | 2 | 295 | Std. Mean Difference (IV, Random, 95% CI) | ‐0.35 [‐0.58, ‐0.12] |
| 1.5 Anti‐rejection drugs | 1 | 17828 | Std. Mean Difference (IV, Random, 95% CI) | ‐0.04 [‐0.07, ‐0.01] |
