Types of studies

We aimed to include studies that, irrespective of publication status and language, evaluated the diagnostic accuracy of platelet count, spleen length, and platelet count‐to‐spleen length ratio for the diagnosis of oesophageal varices with oesophago‐gastro‐duodenoscopy as the reference standard. We considered studies of cross‐sectional cohort design including people with clinical suspicion of portal hypertension as well as studies of participant‐control design that compared people with oesophageal varices versus matched controls (Colli 2014a). We excluded studies that analysed data only per varix rather than per participant unless participant data were made available by study authors.

Participants

Participants included paediatric or adult patients of any age with chronic liver disease or portal vein thrombosis, irrespective of aetiology, severity of disease, and duration of illness, in whom the presence or absence of varices was confirmed by oesophago‐gastro‐duodenoscopy. The review focused on diagnostic questions related to patients who have not yet suffered gastrointestinal bleeding from oesophageal varices. Patients with a previous surgical portal‐systemic shunt procedure or insertion of a transjugular intrahepatic portal‐systemic shunt (TIPS), previous ligation or sclerotherapy of oesophageal varices, previous history of upper gastrointestinal portal hypertensive bleeding, or previous primary prophylactic therapy of variceal haemorrhage make up a distinct group for whom the diagnosis or natural history of oesophageal varices has been modified. These patients were not the focus of this review, hence we excluded studies that included such patients unless investigators presented data in such a way as to allow this patient group to be isolated from other included patients.

Index tests

1. Platelet count is obtained from a complete blood count, a readily available automated clinical test. A platelet count cut‐off value less than 150,000/mm³ is considered thrombocytopenia.

2. Spleen length is usually obtained through evaluation of the patient's abdomen by ultrasound scan (USS). Interobserver agreement when spleen length is determined with USS is considered excellent. For adults, the upper limit of spleen length is 130 mm, beyond which the spleen is generally considered enlarged. Spleen length of 110 mm is regarded as a sensitive cut‐off for exclusion of splenomegaly (Grover 1993). For children, spleen length is expressed as a standard deviation score relative to normal values for both age and sex (spleen length z‐score) (Megremis 2004).

3. Platelet count‐to‐spleen length ratio is a derivative mathematical model shown to increase the accuracy of both non‐invasive tests for the diagnosis of oesophageal varices. The cut‐off value used most often for adults is 909 (n/mm³)/mm. In children, platelet count‐to‐spleen length ratio is calculated using the spleen length z‐score.

Target conditions

The presence of any oesophageal varices (independent of size) was detected by oesophago‐gastro‐duodenoscopy. For secondary analyses, the target condition considered was the presence of oesophageal varices at high risk of bleeding. High‐risk varices were defined as medium or large varices or small varices with red marks, or in patients with decompensated cirrhosis, as assessed by a B‐C Child‐Pugh score (Garcia‐Tsao 2007). Studies will require at least one of two target conditions to be identified: the presence of any oesophageal varices, or the presence of high‐risk varices.

Reference standards

Oesophago‐gastro‐duodenoscopy is the clinical reference standard test for the diagnosis of oesophageal varices in which the presence of varices in the oesophagus is directly observed through the endoscope. The size and appearance of oesophageal varices are graded at the time of endoscopy according to one of the systems described below, and the largest varix identified is used to classify the patient. Severity of cirrhosis, which is the other factor that defines bleeding risk, is assessed by Child‐Pugh score, with three classes ‐ A, B, and C ‐ indicating increasing severity (Pugh 1973). Patients whose largest varix is medium or large or who are included in class B‐C are considered for prophylactic therapy.

1. The Baveno Consensus system differentiates small from large oesophageal varices (de Franchis 1992), defining small oesophageal varices as varices that flatten with insufflation during endoscopy or that minimally protrude into the oesophageal lumen, and large oesophageal varices as varices that protrude into the oesophageal lumen and touch each other, or that fill at least 50% of the oesophageal lumen.

2. The Japanese Research Society for Portal Hypertension used three grades for variceal size (JSPH 1980). Grade 1 varices collapse with insufflation during endoscopy, grade 2 varices do not collapse with insufflation and do not occlude the lumen, and grade 3 varices occlude the lumen. For this review, we will consider grade 2 as equivalent to medium, and grade 3 as large.

3. The Japanese classification was revised by the Italian Liver Cirrhosis Project (ILCP) Group (Pagliaro 1988; Zoli 1996), which describes variceal size as the percentage of the radius of the oesophageal lumen that is occupied by the largest varix. A small or grade 1 varix is said to occupy less than 25%, a medium or grade 2 varix occupies 25% to 50%, and a large or grade 3 varix occupies greater than 50% of the radius of the lumen of the oesophagus.

4. The Cales criteria define varices as small if they flatten with insufflation during endoscopy, medium if they do not flatten with insufflation, and large if they do not flatten with insufflation during endoscopy and are confluent (Cales 1990).

5. We will include studies applying alternate classifications if adequately described and logically defined.

Red marks are usually noted as present or absent and may be described according to different classifications. Even small varices showing red marks are classified as ’at high risk of bleeding’.

The interval between index tests and oesophago‐gastro‐duodenoscopy has to be less than 3 months to avoid possible evolution of the target condition. When a study reported longer time intervals, we included the study but considered it to be at risk of bias. Clinically, patients with medium or large oesophageal varices or with red marks are at greatest risk of haemorrhage; therefore, we confined secondary analyses to two subgroups: patients with no varices and small varices compared with patients at high risk.

Electronic searches

We searched the Cochrane Hepato‐Biliary Group Controlled Trials Register (Gluud 2016), the Cochrane Hepato‐Biliary Group Diagnostic Test of Accuracy Studies Register (Gluud 2016), the Cochrane Library, MEDLINE (OvidSP), Embase (OvidSP), and Science Citation Index ‐ Expanded (Web of Science) (Royle 2003). We have presented in Appendix 1 search strategies along with time spans of the searches. .We applied no language or document‐type restrictions.

Searching other resources

We identified additional references by manually searching the references of articles retrieved from computerised databases and relevant review articles. We sought information on unpublished studies by contacting experts in the field. In addition, we handsearched abstract books from meetings of the American Association for the Study of Liver Diseases (AASLD) and the European Association for the Study of the Liver (EASL) held over the past 10 years.

Selection of studies

We retrieved publications if they were potentially eligible for inclusion on the basis of abstract review, or if they were relevant review articles for a manual reference search. Two review authors independently reviewed publications for eligibility. To determine eligibility, we assessed each publication to determine whether participants met the inclusion criteria detailed above. We included abstracts only if they provided sufficient data for analysis. We resolved disagreements by consensus.

Data extraction and management

Review authors, working in pairs (JCG and JY or AC and GC), completed a data extraction form for each included study. AC and GC completed extraction forms for studies retrieved during the last search (from 2009 to 2016). Each review author independently retrieved study data. In cases of discordance, we reached consensus through discussion.

We retrieved the following data.

1. General information: title, journal, year, publication status, and study design (prospective vs retrospective).

2. Sample size: number of participants meeting the criteria and total number of participants screened.

3. Baseline characteristics: baseline diagnosis, age, sex, race, and disease severity, and medications used concurrently. We considered severity of liver disease among the studied population by using the Child‐Pugh score (Pugh 1973) and the model for end‐stage liver disease (MELD) in adults (Kamath 2001), and by using the Child‐Pugh score and paediatric end‐stage liver disease (PELD) scores in children (McDiarmid 2002).

4. We reported index tests with all cut‐off values.

5. We used the following as clinical reference standard tests: variceal size, type of classification used, number of endoscopists, and handling of interobserver error on oesophago‐gastro‐duodenoscopy.

6. Numbers of true positive (TP), true negative (TN), false positive (FP), and false negative (FN) findings. We extracted these data for each presented cut‐off value and for the two target conditions.

We summarised data from each study in 2 × 2 tables (FP, FN, TP, TN) according to the two target conditions and entered the data into Review Manager 5 software.

Assessment of methodological quality

Two review authors independently assessed the risk of bias of included studies using QUADAS‐2 (revised tool for quality assessment of diagnostic accuracy studies) domains (Whiting 2011). In cases of discordance, we reached a consensus through discussion. We adopted the domains in Appendix 2 to address aspects of study quality involving the participant spectrum, index tests, target conditions, reference standards, and flow and timing. We did not plan to consider blinding of the index test to results of the reference standard for cases in which platelet count is obtained by an automated counter. We classified a study as having high risk of bias if we judged study to have high risk of bias or unclear risk of bias in at least one of the domains of QUADAS‐2.

Statistical analysis and data synthesis

We carried out statistical analyses according to recommendations provided in the Cochrane Handbook for Systematic Reviews of Diagnostic Test Accuracy (DTA Handbook 2010).

We built 2 × 2 tables (TP, TN, FP, FN) for each primary study for the three index tests for the two target conditions (any varices and high‐risk varices). We considered studies with adult participants and studies with paediatric participants separately, as we retrieved only studies that included only adult or paediatric participants.

For all combinations of index test/target condition/participants, we followed the following strategy of analysis. First, we performed a graphical descriptive analysis of the included studies: We reported forest plots (sensitivity and specificity separately, with their 95% confidence intervals (CIs)), and we provided a graphical representation of studies in the receiver operating characteristic (ROC) space (sensitivity against 1 ‐ specificity). Second, we performed a meta‐analysis. When primary studies reported accuracy estimates of an index test using different cut‐off points, we used the hierarchical summary ROC model (HSROC) to pool data (sensitivities and specificities) and to plot a summary ROC (SROC) curve (Rutter 2001). When considering studies with a common cut‐off value, we used the bivariate model and provided estimates of summary sensitivity and specificity. We used pooled estimates obtained from the fitted models to calculate summary estimates of positive and negative likelihood ratios (LR+ and LR‐, respectively).

For primary studies that reported accuracy results for more than one cut‐off point, we reported sensitivities and specificities for all cut‐off points, but we used a single cut‐off point for each study in HSROC (or bivariate) analysis.

We made pair‐wise comparisons between tests by adding a covariate for the index test to the HSROC (for comparisons of SROC curves) or bivariate (for comparisons of sensitivity and specificity at fixed cut‐off value) model. We assessed the significance of differences in test accuracy by using the log‐likelihood ratio test for comparison of models with and without the index test covariate term. We performed both indirect and direct comparisons, if sufficient data were available.

We considered P values less than 0.05 as two‐sided and statistically significant.

We performed all statistical analyses using SAS statistical software, release 9.4 (SAS Institute Inc., Cary, NC, USA) and macro METADAS (DTA Handbook 2010).

Investigations of heterogeneity

We investigated effects of the following predefined sources of heterogeneity.

1. Chronic liver disease compared with portal vein thrombosis.

2. Prevalence of oesophageal varices in the study group (≥ 50% versus < 50% for any varices; > 25% versus ≤ 25% for high‐risk varices).

3. Severity of liver disease Child A (> 50% versus ≤ 50%).

4. Different aetiologies (HCV‐associated cirrhosis versus all aetiologies),

by adding covariates to the bivariate or to the HSROC. We assessed the statistical significance of the covariate effect by using the log‐likelihood ratio test for comparison of models with and without the covariate term.

To limit the number of statistical analyses, we investigated sources of heterogeneity by considering only studies with the cut‐off value defined in the "Index test" section.

Sensitivity analyses

We attempted to assess effects of risk of bias of included studies on diagnostic accuracy by performing a sensitivity analysisfrom which we excluded studies with the following characteristics.

1. Studies classified at high risk of bias. We classified a study as having high risk of bias if we judged study to have high risk of bias or unclear risk of bias in at least one of the domains of QUADAS‐2 (Appendix 2). In addition, we identified the two following signalling questions as most relevant, and we decided to assess them in separate sensitivity analyses.

   1. "Was a case‐control design avoided?"

   2. "If a threshold was used, was it prespecified?"

2. Studies published only in abstract/letter form.

To limit the number of statistical analyses, we performed sensitivity analyses by considering only studies with the cut‐off value defined in the "Index test" section.