Design

All the studies we included were RCTs, comprising one cluster‐RCT (McCorkle 2015), one cluster‐randomised crossover trial (Ma 2019) and eight fast‐track RCTs (Bajwah 2015; Bakitas 2015; Edmonds 2010; Farquhar 2014; Farquhar 2016; Higginson 2009; Higginson 2014; McWhinney 1994). The remaining 32 RCTs had a parallel design.

Sample sizes

Sample sizes in included studies ranged from 30 to 621 participants. The length of recruitment in included studies varied between 10 months and 50 months. In total, we included data from studies involving 7779 participants (6678 adults with advanced illness and 1101 unpaid caregivers/family members). Thirty‐three studies had power calculations (details in 'Characteristics of included studies'). Nine studies were powered on quality of life only (Bekelman 2018; El‐Jawahri 2016; Franciosi 2019; Groenvold 2017; Rogers 2017; Tattersall 2014; Temel 2010; Temel 2017; Vanbutsele 2018). Ma 2019 was powered on proportion of patients transitioning to do‐not‐resuscitate and do‐not‐intubate (DNR/DNI). In addition to quality of life, Bakitas 2015 also performed calculations on depression, Solari 2018 on symptom burden, O'Riordan 2019 on pain, while Bakitas 2009 and Sidebottom 2015 included depression and symptom burden. Farquhar 2014 and Farquhar 2016 were powered on distress due to breathlessness, Brannstrom 2014 on symptom burden, Brumley 2007 on cost, Carson 2016 on depression and anxiety, Grudzen 2016 on time to palliative care, Janssens 2019 on hospital admission, Rodin 2019 on traumatic stress symptoms, Bajwah 2015, Edmonds 2010 and Higginson 2009 on the Palliative care Outcome Scale (POS), Lowther 2015 on the African Palliative care Outcome Scale (APOS), Higginson 2014 on Chronic Respiratory Disease Questionnaire (CRDQ) mastery domain, Hopp 2016 and Ozcelik 2014 on palliative outcomes and palliative care service respectively, McWhinney 1994 on pain and nausea and Woo 2019 on pain and depression.

Eight studies were well‐powered at recruitment and also at the primary point of analyses (Carson 2016; Edmonds 2010; Farquhar 2016; Higginson 2014; Ma 2019; Ozcelik 2014; Solari 2018; Temel 2017). Fourteen studies were underpowered at recruitment stage (i.e. participants enrolled) by three participants (Brumley 2007; Groenvold 2017; Hopp 2016), four (Grudzen 2016), eight (Rodin 2019), 19 (Nottelmann 2018), 25 (O'Riordan 2019), 30 (Tattersall 2014), 50 (Rogers 2017), 74 (McWhinney 1994), 78 (Bakitas 2009), 111 (Janssens 2019), 153 (Bakitas 2015) and 268 (Sidebottom 2015). In one of the underpowered studies (Rogers 2017), the data and safety monitoring board in consultation with the sponsoring agency recommended a sample size reduction due to enrollment rates, a mortality rate that was lower than predicted and observed outcomes differences at the intermediate time point. Reasons provided for underpowered studies included slower than anticipated accrual, resource constraints, early deaths, problems with recruitment and low compliance rate for completion of questionnaires. The remaining 11 studies included the numbers that they had planned to recruit but dropped below the required numbers by the first time point of analyses (i.e. following baseline assessment and after receiving the intervention or control). These studies were underpowered by two participants (Brannstrom 2014), three participants (El‐Jawahri 2016), five participants each (Bajwah 2015; Higginson 2009), six participants each (Lowther 2015; Farquhar 2014), 13 participants (Temel 2010), 22 participants (Vanbutsele 2018), 29 participants (Franciosi 2019), 60 participants (Woo 2019) and 70 participants (Bekelman 2018). Nine studies did not report any power calculation (Ahronheim 2000; Cheung 2010; Gade 2008; Jingfen 2017; Kane 1984; Mendoza‐Galindo 2018 (abstract only); McCaffrey 2013; McCorkle 2015; Wallen 2012) (see Figure 2 for a graphical representation of the power of included studies at recruitment and follow‐up). Overall, 14 studies examined post‐intervention assessments in more than 100 participants.

Figure 2

---

---

A figure describing the power of included studies at recruitment and follow‐up

Setting

Nineteen studies were carried out in USA. Six studies took place in the UK (Bajwah 2015; Edmonds 2010; Farquhar 2014; Farquhar 2016; Higginson 2009; Higginson 2014), and three studies occurred in Australia (Cheung 2010; McCaffrey 2013; Tattersall 2014). One study was conducted in Sweden (Brannstrom 2014), two in Denmark (Groenvold 2017; Nottelmann 2018), one in Switzerland (Janssens 2019), one in Belgium (Vanbutsele 2018), two in Italy (Franciosi 2019; Solari 2018), and one in Turkey (Ozcelik 2014). McWhinney 1994 and Rodin 2019 were carried out in Canada, while Woo 2019 was undertaken in South Korea. Lowther 2015 took place in Kenya and Jingfen 2017 was conducted in China. Mendoza‐Galindo 2018 (abstract only) was carried out in Mexico.

Thirty studies recruited from hospital settings. Three of these studies recruited from intensive care units (ICU) (Carson 2016; Cheung 2010; Ma 2019). Of these 30 studies, Ahronheim 2000 recruited patients with advanced dementia from Mount Sinai Hospital, Bajwah 2015 recruited from a specialist interstitial lung disease centre, Janssens 2019 from patients followed by Geneva University Hospitals on long‐term oxygen therapy (LTOT) and/or home non‐invasive ventilation (NIV) as well as those hospitalised for acute exacerbation of Chronic Obstructive Oulmonary Disease (COPD) in the general internal medicine and geriatric wards, Lowther 2015 from outpatient HIV clinics in a community hospital, McCorkle 2015 from disease‐specific multidisciplinary clinics at a cancer hospital, O'Riordan 2019 from new inpatient admissions to the medicine and cardiology services, Solari 2018 from three Italian multiple sclerosis centres and Franciosi 2019 from outpatient and inpatient settings at five Italian cancer centres. Seven studies recruited from oncology centres or clinics (Groenvold 2017; Rodin 2019; Tattersall 2014; Temel 2010; Temel 2017; Vanbutsele 2018; Woo 2019). Bakitas 2009 and Bakitas 2015 recruited from oncology clinics of a cancer centre and affiliated outreach clinics and the Veterans Affairs Medical Centre (VAMC).

Eleven studies recruited from primary care and/or secondary care. For example, Gade 2008 recruited from community medical services and inpatient units, while McWhinney 1994 recruited through family physicians and home care nurses. Brumley 2007 received referrals from discharge planners, primary care physicians and other specialty physicians, whereas Rogers 2017 enrolled both hospitalised patients and recently discharged patients who were at high risk of rehospitalisation. Higginson 2009 received referrals from local health and social care professionals. Edmonds 2010 received referrals from health and social care professionals and, in a few instances, through voluntary organisations and self‐referral.

Mendoza‐Galindo 2018 (abstract only) did not present the setting where recruitment took place..

Participants

Twenty‐one studies were carried out with patients who had severe/advanced cancer or their unpaid caregivers/family members or both (Bakitas 2009; Bakitas 2015; El‐Jawahri 2016; Farquhar 2014; Franciosi 2019; Groenvold 2017; Grudzen 2016; Jingfen 2017; Kane 1984; McCorkle 2015; McWhinney 1994; Mendoza‐Galindo 2018 (abstract only); Nottelmann 2018; Ozcelik 2014; Rodin 2019; Tattersall 2014; Temel 2010; Temel 2017; Vanbutsele 2018; Wallen 2012; Woo 2019). A range of cancers were included in these studies comprising solid and non‐solid tumour cancers. Seven studies involved both cancer and non‐cancer populations (mixed populations) (Brumley 2007; Carson 2016; Cheung 2010; Gade 2008; Higginson 2014; Ma 2019; McCaffrey 2013), while the remaining 14 studies had only non‐cancer populations. The non‐cancer populations were those with interstitial lung disease (Bajwah 2015), heart failure (Bekelman 2018; Brannstrom 2014; Hopp 2016; O'Riordan 2019; Rogers 2017; Sidebottom 2015), HIV (Lowther 2015), dementia (Ahronheim 2000), multiple sclerosis (Edmonds 2010; Higginson 2009; Solari 2018), COPD (Janssens 2019) and a combination of COPD (83%) and other non‐malignant diseases (Farquhar 2016). Two studies were with rural populations (Bakitas 2009; Bakitas 2015), while Hopp 2016 was with a predominantly African‐American population (92%). Thirty‐five studies were conducted or first published from 2010 onwards, with 89% taking place within the last six years (see Characteristics of included studies for details).

Mean/median age ranged from 38.3 to 85.6 years. About the same number of males and females were included in most studies. However, five studies had between 69% and 82% females (Ahronheim 2000; Edmonds 2010; Higginson 2009; Lowther 2015; Ozcelik 2014), whereas nine studies had 60% to 98% males (Bajwah 2015; Bakitas 2009; Bekelman 2018; Brannstrom 2014; Farquhar 2016; Franciosi 2019; Kane 1984; Rodin 2019; Vanbutsele 2018). Ahronheim 2000 had the highest percentage of females (82%). Kane 1984 who recruited at a Veterans Administration hospital included predominantly male veterans. Wallen 2012 did not provide the gender distribution in their population. Unpaid caregivers/family members included in studies tended to be mainly females. Nine of the 16 studies involving unpaid caregivers/families described one or more of their characteristics: they were mainly spouses and women, and had a median/mean age ranging from 51 to 65.6 years. In five studies, between 16% and 43% of patients lived alone (Farquhar 2014; Farquhar 2016; Higginson 2009; McCorkle 2015; Vanbutsele 2018).

Sixteen studies had survival as an inclusion criterion. Life expectancy specified in these studies ranged from > 72 hours to 24 months. Eight studies specifically stated that they included newly diagnosed patients (Bakitas 2015; Franciosi 2019; McCorkle 2015; Nottelmann 2018; Rodin 2019; Temel 2010; Temel 2017; Woo 2019). Exclusion criteria included the presence of severe mental illness (Bakitas 2009; Bakitas 2015; Hopp 2016; Jingfen 2017; Temel 2017), and palliative care/hospice involvement previously or at present/request for palliative care involvement (Bajwah 2015; Carson 2016; Cheung 2010; Franciosi 2019; Grudzen 2016; Ma 2019; Nottelmann 2018; Rodin 2019; Sidebottom 2015; Solari 2018; Tattersall 2014; Temel 2010; Temel 2017; Vanbutsele 2018). In three studies, patients without surrogate decision‐makers were excluded (Carson 2016; Cheung 2010; Solari 2018), while Gade 2008 excluded patients if they had impaired cognitive status and no surrogate. Janssens 2019 and Rodin 2019 excluded patients with moderate or severe cognitive impairment.

Intervention

Taxonomy of the components of HSPC

We assessed the components of HSPC in the studies included in this review using the principles and domains of palliative care highlighted by Zimmermann 2019. Zimmermann 2019 developed a conceptual framework highlighting the domains and principles of team‐based outpatient early palliative care for patients with cancer. This framework is based on palliative care theory (Doyle 1998; WHO 2002; Zimmermann 2004; Zimmermann 2012), review of previous palliative care interventions (Zimmermann 2008) and practice guidelines (Cancer Care Ontario 2016; NCCN 2016). This framework was chosen above others such as the Holistic Common Assessment (National End of Life Care Programme 2010), which is used for comprehensive palliative care assessment, because the essential elements of the framework are consistent with the need for early provision of palliative care in collaboration with the multidisciplinary team, and also because it is based on the needs of the patient and their family, rather than on prognosis.

The four domains are coping and support, decision‐making, symptom control and future planning, while the four principles are that care is flexible, attentive, patient‐ and family‐centred.

Components of HSPC in studies that either included certified experts in palliative care or those described as palliative care clinicians

Thirty‐one studies either included certified experts in palliative care or those described as palliative care clinicians. Eight studies were only patient‐centred (Brannstrom 2014; Rodin 2019; Rogers 2017; Sidebottom 2015; Tattersall 2014; Temel 2017; Vanbutsele 2018; Wallen 2012), while Carson 2016 was only family‐centred because the intervention was a palliative care‐led meeting for families of patients in the medical intensive care unit. Twenty‐two studies were both patient‐centred and family‐centred (Bajwah 2015; Bakitas 2009; Bakitas 2015; Bekelman 2018; Brumley 2007; Edmonds 2010; El‐Jawahri 2016; Farquhar 2014; Farquhar 2016; Franciosi 2019; Gade 2008; Higginson 2009; Higginson 2014; Janssens 2019; Kane 1984; Lowther 2015; Ma 2019; McCorkle 2015; McWhinney 1994; Nottelmann 2018; Solari 2018; Temel 2010). For instance, the HSPC intervention in Bajwah 2015 was individualised to each patient and carer, while Vanbutsele 2018 described the use of semi‐structured monthly consultations by palliative care nurses that allowed for individualised care. Bekelman 2018 described collaboration between patients and the nurse as they both agreed on the symptom to focus on.

Palliative care in all 31 studies except Kane 1984 involved provision of care that was flexible and attentive to the needs of patients and/or their families as they allowed for the involvement of other members of the healthcare team in order to address these needs.

We mapped the 31 studies to the four domains highlighted above. We added care co‐ordination as an additional domain because the need for co‐ordinated care for those with advanced disease is not always delivered and this can result in increased hospitalisations and suboptimal clinical outcomes (Higginson 2003; Walsh 2011) (see Table 1 under Additional tables for the domains covered in the studies and Figure 3 for the percentage of studies assessing different domains).

Open in table viewer

Table 1. Taxonomy of the components of hospital‐based specialist palliative care in studies that either included certified experts in palliative care or those described as palliative care clinicians

Author	Symptom control (e.g. assess symptoms, prescribing of medications)	Decision‐making (e.g. enquire about goals of care)	Future planning (e.g. advance care planning)	Coping and support (e.g. emotional and practical support)	Care co‐ordination (e.g. helping with co‐ordinating care)
Bajwah 2015	Yes	Yes	Yes	Yes	Yes
Bakitas 2009	Yes	Yes	Yes	Yes	Yes
Bakitas 2015	Yes	Yes	Yes	Yes	Yes
Bekelman 2018	Yes	Yes	No	Yes	Yes
Brannstrom 2014	Yes	Yes	No	Yes	Yes
Brumley 2007	Yes	Yes	Yes	Yes	Yes
Carson 2016	No	Yes	No	Yes	No
Edmonds 2010	Yes	Yes	Yes	Yes	Yes
El‐Jawahri 2016	Yes	No	No	Yes	No
Farquhar 2014	Yes	Yes	Yes	Yes	No
Farquhar 2016	Yes	Yes	Yes	Yes	No
Franciosi 2019	Yes	Yes	No	Yes	Yes
Gade 2008	Yes	Yes	Yes	Yes	No
Higginson 2009	Yes	No	Yes	Yes	Yes
Higginson 2014	Yes	Yes	Yes	Yes	Yes
Janssens 2019	Yes	Yes	Yes	Yes	Yes
Kane 1984	Yes	No	Yes	Yes	No
Lowther 2015	Yes	Yes	Yes	Yes	No
Ma 2019	Yes	Yes	No	Yes	Yes
McCorkle 2015	Yes	Yes	No	Yes	Yes
McWhinney 1994	Unclear	Unclear	Unclear	Yes	Unclear
Nottelmann 2018	Yes	Yes	Yes	Yes	Yes
Rodin 2019	Yes	No	No	Yes	No
Rogers 2017	Yes	Yes	Yes	Yes	Yes
Sidebottom 2015	Yes	Yes	Yes	Yes	Yes
Solari 2018	Unclear	Unclear	Unclear	Yes	Unclear
Tattersall 2014	Yes	No	No	Yes	No
Temel 2010	Yes	Yes	No	Yes	Yes
Temel 2017	Yes	Yes	No	Yes	Yes
Vanbutsele 2018	Yes	Yes	No	Yes	Yes
Wallen 2012	Yes	No	No	Yes	No

Figure 3

---

---

A figure showing the domains of HSPC in the studies that either included certified experts in palliative care or those described as palliative care clinicians

Symptom control

This involved assessment and management of symptoms. Twenty‐eight studies highlighted that the HSPC intervention included symptom or needs assessment and management. In two studies, this was unclear (McWhinney 1994; Solari 2018), while it appeared that Carson 2016 did not address this domain.

Decision‐making

This domain involved assessing patient and their family's understanding of illness, assessing individual and cultural values/beliefs or assessing goals of care and regularly reviewing them. Twenty‐three studies included HSPC that addressed one or more aspects of decision‐making. One study included HSPC that did not focus on decision‐making as the intervention was targeted at managing patients’ physical and psychological symptoms during hospitalisation (El‐Jawahri 2016). A further five studies of HSPC did not involve this domain (Higginson 2009; Kane 1984; Rodin 2019; Tattersall 2014; Wallen 2012). In two studies, this was unclear (McWhinney 1994; Solari 2018).

Future planning

Future planning involved discussing concerns and preferences for end‐of‐life care, making a will, power of attorney and decisions about resuscitation. Half of the studies (n = 16) included HSPC that involved planning for the future, while this was unclear in two studies (McWhinney 1994; Solari 2018). The remaining 13 studies did not include this domain in delivery of HSPC (Bekelman 2018; Brannstrom 2014; Carson 2016; El‐Jawahri 2016; Franciosi 2019; Ma 2019; McCorkle 2015; Rodin 2019; Tattersall 2014; Temel 2010; Temel 2017; Vanbutsele 2018; Wallen 2012).

Coping and support

This involved establishing a therapeutic relationship, facilitating coping with advanced illness and spiritual support, providing emotional and practical support, addressing family needs and bereavement care.

All 31 studies involved HSPC that had one or more elements of this domain. In particular, three studies included HSPC that specifically highlighted bereavement care or involved a bereavement co‐ordinator as needed (Bakitas 2009; Brumley 2007; Higginson 2009). Bakitas 2009 provided a bereavement follow‐up call to the unpaid caregiver as part of the HSPC intervention, while Higginson 2009 described HSPC that provided bereavement support when needed. Brumley 2007 described HSPC that included a bereavement co‐ordinator, as needed. Furthermore, Bekelman 2018 included a topic on grief and loss as part of the counselling session in their HSPC intervention.

In addition to the areas described above, we assessed the provision of spiritual care. Thirteen studies include HSPC provided spiritual care or support (Bajwah 2015; Brannstrom 2014; Brumley 2007; Carson 2016; Higginson 2014; Janssens 2019; Kane 1984; Lowther 2015; Nottelmann 2018; Rogers 2017; Sidebottom 2015; Vanbutsele 2018; Wallen 2012).

Care co‐ordination

Although Zimmermann 2019 did not include care co‐ordination as a domain in their conceptual framework, we decided to include this domain. Over half of the studies (n = 19) involved HSPC that provided care co‐ordination (Bajwah 2015; Bakitas 2009; Bakitas 2015; Bekelman 2018; Brannstrom 2014; Brumley 2007; Edmonds 2010; Franciosi 2019; Higginson 2009; Higginson 2014; Janssens 2019; Ma 2019; McCorkle 2015; Nottelmann 2018; Rogers 2017; Sidebottom 2015; Temel 2010; Temel 2017; Vanbutsele 2018), while this was unclear in two studies (McWhinney 1994; Solari 2018). In 10 studies, it appeared that the HSPC intervention did not include this domain.

The main domains of care in the HSPC intervention in the 31 studies were symptom control, coping and support, and decision‐making. At least half of the studies included HSPC that provided care co‐ordination and future planning. Besides McWhinney 1994 and Solari 2018, the remaining studies addressed at least two domains.

Components of HSPC in studies that were unclear about palliative care training

Eleven studies were unclear about the palliative care training of those who delivered the HSPC intervention (Ahronheim 2000; Cheung 2010; Groenvold 2017; Grudzen 2016; Hopp 2016; Jingfen 2017; McCaffrey 2013; Mendoza‐Galindo 2018 (abstract only); O'Riordan 2019; Ozcelik 2014; Woo 2019) (see Table 2 under Additional tables for the domains covered in the studies and Figure 4 for the percentage of studies assessing different domains). Four of these studies were only patient‐centred (Groenvold 2017; Hopp 2016; O'Riordan 2019; Woo 2019), while Ahronheim 2000 was only family‐centred. Three studies were both patient‐ and family‐centred (Grudzen 2016; Jingfen 2017; Ozcelik 2014). In three studies, this was unclear (Cheung 2010; McCaffrey 2013; Mendoza‐Galindo 2018 (abstract only)).

Open in table viewer

Table 2. Taxonomy of the components of hospital‐based specialist palliative care in studies that were unclear about training in palliative care

Author	Symptom control (e.g. assess symptoms, prescribing of medications)	Decision‐making (e.g. enquire about goals of care)	Future planning (e.g. advance care planning)	Coping and support (e.g. emotional and practical support)	Care co‐ordination (e.g. helping with co‐ordinating care)
Ahronheim 2000	Yes	No	Yes	Yes	No
Cheung 2010	Unclear	Unclear	Unclear	Unclear	Unclear
Groenvold 2017	Unclear	Unclear	Unclear	Unclear	Unclear
Grudzen 2016	Yes	Yes	Yes	Yes	No
Hopp 2016	Yes	Yes	Yes	Yes	No
Jingfen 2017	Yes	Yes	No	Yes	No
McCaffrey 2013	Unclear	Unclear	Unclear	Unclear	Yes
Mendoza‐Galindo 2018 (abstract only)	Yes	No	No	Yes	No
O'Riordan 2019	Yes	No	Yes	Yes	No
Ozcelik 2014	Yes	No	Yes	Yes	No
Woo 2019	Yes	No	No	Yes	No

Figure 4

---

---

A figure showing the domains of HSPC in studies that were unclear about palliative care training

Palliative care provision was flexible in all the 11 studies, with the involvement of the multidisciplinary team as required to address the needs of patients and/or their families. In 10 studies, the palliative care providers were attentive to the needs of patients and their families, while this was unclear in Mendoza‐Galindo 2018 (abstract only). Mendoza‐Galindo 2018 (abstract only) only reported that the intervention was provided by a palliative team, which included psychological, nutritional and symptom support.

We assessed the domains of HSPC included in these studies as follows:

Symptom control

Eight studies highlighted the issue that the HSPC intervention included symptom or needs assessment and management. In three studies, this was unclear (Cheung 2010; Groenvold 2017; McCaffrey 2013).

Decision‐making

Three studies involved one or more aspects of decision‐making (Grudzen 2016; Hopp 2016; Jingfen 2017), while this was unclear in three studies (Cheung 2010; Groenvold 2017; McCaffrey 2013). It appeared five studies did not involve this domain (Ahronheim 2000; Mendoza‐Galindo 2018 (abstract only); O'Riordan 2019; Ozcelik 2014; Woo 2019).

Future planning

Five studies involved planning for the future (Ahronheim 2000; Grudzen 2016; Hopp 2016; O'Riordan 2019; Ozcelik 2014), while this was unclear in three studies (Cheung 2010; Groenvold 2017; McCaffrey 2013). Three studies did not include this domain (Jingfen 2017; Mendoza‐Galindo 2018 (abstract only); Woo 2019).

Coping and support

Eight studies involved one or more elements of this domain, while three studies were unclear (Cheung 2010; Groenvold 2017; McCaffrey 2013). O'Riordan 2019 further involved the provision of spiritual care.

Care co‐ordination

Only McCaffrey 2013 involved care co‐ordination, while eight studies did not. It was unclear whether two studies included this domain (Cheung 2010; Groenvold 2017).

The main domains of care in the HSPC intervention in studies with unclear training were symptom control, coping and support, and future planning. Very few studies involved decision‐making and care co‐ordination. Besides three studies where the domains were unclear, the remaining eight studies addressed at least two domains.

When compared to studies that included experts or those described as palliative care clinicians, studies with unclear palliative care training often did not include decision‐making and care co‐ordination. There was also less focus on symptom control, and coping and support in studies with unclear palliative care training. Both groups were similar in the extent to which they focussed on future planning.

Controls

HSPC was compared with usual care. Overall, there was poor description of usual care in most studies with no information or very little information provided. For example, Ahronheim 2000 only stated that the control group was treated by the primary care team without palliative care input, and Cheung 2010 stated that control group received usual ICU care without palliative care consultation. Among studies providing some level of detail on usual care, it appeared usual care was varied, probably reflecting the local context as well as differences in health systems. For example, in the Kenyan study by Lowther 2015, those in the usual care group received care from nurses without experience in palliative care from the HIV clinic, consisting of monthly clinical assessments once antiretroviral therapy (ART) was established. In the Swiss study by Janssens 2019, patients receiving usual care had no contact with the palliative care team. Specialised nurses provided regular home visits to patients under long‐term oxygen therapy (LTOT) and/or home non‐invasive ventilation (NIV). In the Belgian study by Vanbutsele 2018, usual oncology care in all the participating departments was provided by a multidisciplinary team, including oncologists, other medical specialists, social workers, psychologists, dieticians and specialist nurses. All patients with advanced cancer usually have an introductory consultation with a specialist nurse trained in oncological care, a dietician, and a psychologist at the start of their treatment. Follow‐up consultations were at the patient's discretion. The palliative care team was only involved on demand and often late in the disease trajectory, and their services were not systematically offered to all patients from oncology departments. Usual care in the South Korean study by Woo 2019 comprised anticancer and symptom control treatments and consultation with psychiatric and pain specialists. In Bajwah 2015, a UK study, the control group remained under interstitial lung disease (ILD) specialist care which included input from ILD physicians, ILD clinical nurse specialist, occupational therapists, physiotherapists and oxygen assessment and treatment services. All patients were also able to access inpatient ILD treatment, as needed. The control group received the intervention after four weeks. In Higginson 2014, the control group received usual care services according to UK guidance. After six weeks, the control group was offered the intervention. Similarly, in Solari 2018, usual care involved health and social services provided by the Italian National Health Service and dyads were offered the intervention at the end of the study.

In 20 studies, usual care included involvement of palliative care professionals if needed (Bajwah 2015; Bakitas 2009; Bakitas 2015; Bekelman 2018; Carson 2016; El‐Jawahri 2016; Franciosi 2019; Groenvold 2017; Grudzen 2016; Ma 2019; McCaffrey 2013; McWhinney 1994; Rodin 2019; Rogers 2017; Tattersall 2014; Temel 2010; Temel 2017; Vanbutsele 2018; Wallen 2012; Woo 2019), and, in one study, usual care incorporated hospice care (Brumley 2007). Wallen 2012 reported that the usual care group was permitted to cross over to the intervention arm if standard pain and symptom management was inadequate to meet their needs.

Outcomes

Our primary outcomes of patient health‐related quality of life (HRQoL) and their symptom burden (assessed using generalised measures) were assessed by 10 and six studies reporting adjusted endpoint values, respectively. Of the 10 studies assessing patient HRQoL, nine were with cancer populations, and one with non‐cancer populations. Nine of the 10 studies involved early palliative care (Bakitas 2009; Bakitas 2015; El‐Jawahri 2016; McCorkle 2015; Rodin 2019; Tattersall 2014; Temel 2010; Temel 2017; Vanbutsele 2018). All six studies that reported symptom burden were with cancer populations and they involved early palliative care (Bakitas 2009; Bakitas 2015; El‐Jawahri 2016; Rodin 2019; Tattersall 2014; Temel 2010).

Other patient outcomes reported included individual symptoms (pain, anxiety, depression, post‐traumatic stress disorder, breathlessness, fatigue, nausea/vomiting, appetite loss, sleep disturbance), traumatic stress symptoms, mortality (death at home, hospital and ICU), survival, advanced care planning, functional independence, achieving preferred place of care or death, satisfaction with care, physical function, psychological, social and spiritual well‐being, nutrition, and cognitive status.

Unpaid caregiver outcomes assessed in studies included satisfaction with care, symptom control (e.g. anxiety, depression), HRQoL, burden, coping, distress with patients' symptoms, and grief.

Economic data

We included 31 studies in the economic component of this review as they compared the resource use and/or costs/cost‐effectiveness between HSPC and usual care alongside clinical effectiveness. We restricted the economic component of the review to economic analyses conducted alongside the studies meeting eligibility criteria for the effectiveness component of the review. Of the 31 studies, four studies were full economic evaluations that compared the costs and effects of the intervention and control group between baseline and follow‐up (Farquhar 2014; Farquhar 2016; Higginson 2009; McCaffrey 2013), five partial economic evaluations that compared only costs and outcomes without reporting incremental changes or decision criteria (Brumley 2007; Gade 2008; Higginson 2014; Kane 1984; Temel 2010), and 22 studies reported more limited resource use/cost information.

The studies measured resource use associated with care received in the intervention and the control group. Use of the following resources was assessed: institutional care services use (e.g. emergency department (ED) or accident and emergency (A&E), intensive care unit use, inpatient stay, care in nursing homes (or skilled nursing homes)); outpatient clinic use (e.g. palliative care visits in outpatient settings, consultation with experts in outpatient settings); community care services use (e.g. GP contacts, nurse visits, home care, hospice care at home); unpaid caregiver care; medications and other resource use. Thirteen studies calculated the costs associated with resource utilisation (Brannstrom 2014; Brumley 2007; Farquhar 2014; Farquhar 2016; Gade 2008; Higginson 2009; Higginson 2014; Kane 1984; Ma 2019; McCaffrey 2013; Mendoza‐Galindo 2018 (abstract only); Ozcelik 2014; Temel 2010). Four studies reported the results of cost‐effectiveness analysis using relevant outcome measures (palliative outcome, unpaid caregiver burden, quality‐adjusted life years (QALYs)) (Farquhar 2014; Farquhar 2016; Higginson 2009; McCaffrey 2013), and hospital costs or total costs. Results of cost‐effectiveness analyses were reported by incremental cost‐effectiveness ratios (ICERs) and/or costs per QALY (point estimates or cost‐effectiveness planes). The four studies reported ICERs, cost/QALY, or cost‐effectiveness planes from cost‐effectiveness analysis.

Random sequence generation

Twenty‐seven studies were randomised and provided adequate description of the sequence generation process. We therefore judged them to be at low risk of bias. However, we judged 15 studies to be at unclear risk of bias due to insufficient description of the sequence generation process (Ahronheim 2000; Grudzen 2016; Hopp 2016; Kane 1984; Lowther 2015; Ma 2019; McCaffrey 2013; Mendoza‐Galindo 2018 (abstract only); O'Riordan 2019; Ozcelik 2014; Rogers 2017; Sidebottom 2015; Temel 2010; Wallen 2012; Woo 2019).

Allocation concealment

Authors of 21 studies did not provide adequate information on how they concealed the allocation and so we judged them to be at unclear risk of bias (Ahronheim 2000; Bakitas 2009; Bakitas 2015; Brannstrom 2014; Brumley 2007; El‐Jawahri 2016; Gade 2008; Hopp 2016; Jingfen 2017; Kane 1984; Lowther 2015; McCaffrey 2013; McCorkle 2015; Mendoza‐Galindo 2018 (abstract only); O'Riordan 2019; Ozcelik 2014; Rogers 2017; Sidebottom 2015; Temel 2010; Wallen 2012; Woo 2019). We judged 21 studies as having a low risk of bias because the methods used to conceal the allocation sequence were described.

Blinding of participants and personnel (subjective outcomes)

None of the studies that reported on subjective outcomes blinded participants. We judged two studies as having an unclear risk of bias because they did not state whether participants and personnel were blinded (Jingfen 2017; Mendoza‐Galindo 2018 (abstract only)), while we gave the remaining 36 studies a high risk of bias because they did not carry out blinding. Four studies did not include subjective outcomes and we therefore did not assess this domain in these studies (Ahronheim 2000; Hopp 2016; Ma 2019; McCaffrey 2013) and we left the domain blank. Generally, in palliative care research, blinding of participants and personnel is not possible or feasible due to the nature of palliative care interventions which involve service provision by a multidisciplinary team (Piggott 2004), and also because of ethical considerations as patients need to be informed about the intervention.

Blinding of participants and personnel (objective outcomes)

We judged 29 studies to be at low risk of bias because we considered that lack of blinding of participants and personnel was unlikely to affect the objective outcomes they assessed. We could not assess this domain in 12 studies because they did not include objective outcomes (Edmonds 2010; Farquhar 2014; Farquhar 2016; Higginson 2009; Jingfen 2017; Lowther 2015; McCorkle 2015; McWhinney 1994; O'Riordan 2019; Ozcelik 2014; Temel 2017; Wallen 2012) and so we therefore left this domain blank. We gave Mendoza‐Galindo 2018 (abstract only) an unclear risk of bias because the authors did not state whether blinding of participants and personnel occurred.

Blinding of outcome assessment (subjective outcomes)

We judged nine studies as having a low risk of bias because they blinded outcome assessors (Bakitas 2015; Bekelman 2018; Brumley 2007; Farquhar 2014; Farquhar 2016; Franciosi 2019; Groenvold 2017; McWhinney 1994; Solari 2018). We assessed 14 studies as having an unclear risk of bias rating (Carson 2016; Gade 2008; Grudzen 2016; Higginson 2014; Jingfen 2017; Kane 1984; McCorkle 2015; Mendoza‐Galindo 2018 (abstract only); Nottelmann 2018; O'Riordan 2019; Ozcelik 2014; Sidebottom 2015; Wallen 2012; Woo 2019) because it was unclear whether outcome assessors were blinded, while we gave 15 studies a high risk of bias rating because they did not blind outcome assessors. Some authors of studies with a high risk of bias stated explicitly that they did not blind outcome assessors (e.g. Vanbutsele 2018), while others stated that they were open‐label or non‐blinded studies (e.g. Bakitas 2009; Janssens 2019; Rodin 2019; Temel 2017).

We could not assess this domain in four studies because they did not include subjective outcomes (Ahronheim 2000; Hopp 2016; Ma 2019; McCaffrey 2013) so we therefore left this domain blank.

Blinding of outcome assessment (objective outcomes)

We assessed 29 studies as having a low risk of bias because they blinded outcome assessors, while we rated two studies as having an unclear risk of bias (Jingfen 2017; Mendoza‐Galindo 2018 (abstract only)) due to lack of clarity on whether outcome assessors were blinded. We could not rate the remaining 11 studies because they did not include objective outcomes (Edmonds 2010; Farquhar 2014; Farquhar 2016; Higginson 2009; Lowther 2015; McCorkle 2015; McWhinney 1994; O'Riordan 2019; Ozcelik 2014; Temel 2017; Wallen 2012) and so we left this domain blank.

BMJ Checklist for authors and peer reviews of economic submissions

The methodological quality of the 13 studies that examined total costs varied across the different areas assessed (see Appendix 8). We assessed methodological quality using the BMJ Checklist for authors and peer reviewers of economic submissions (Drummond 1996). Given that they used different methods and reported on different resources used by patients, we could not pool their data in a meta‐analysis. All the studies were clear about their research question. We considered all the studies to have provided the rationale for choosing the alternatives they compared because they all compared HSPC (or HSPC in addition to usual care) with usual care. However, only eight of them stated the economic importance of the research question. Six studies stated the form of economic evaluation used. The viewpoint of the analysis was stated only in three studies (Higginson 2009; McCaffrey 2013; Sahlen 2016 (linked to Brannstrom 2014)). All studies were clear about the source of effectiveness estimates used. Besides Mendoza‐Galindo 2018 (abstract only), they all provided details on the design and results of their effectiveness study. The primary outcome for the economic evaluation was clearly stated in seven studies (Farquhar 2014; Farquhar 2016; Gade 2008; Higginson 2009; Higginson 2014; McCaffrey 2013; Sahlen 2016 (linked to Brannstrom 2014)). Quantities of resources were not reported separately from their unit costs in four studies (Ma 2019; Mendoza‐Galindo 2018 (abstract only); Ozcelik 2014; Sahlen 2016 (linked to Brannstrom 2014)). In Brumley 2007, this was unclear because the authors described how the costs were derived but did not present the unit costs. Details of currency of price adjustments for inflation or currency conversion were not provided in any of the studies. The relevance of productivity changes to the study question was also not discussed in any of the studies. All studies except Mendoza‐Galindo 2018 (abstract only) stated the time horizon of costs and benefits. They all addressed the research question with conclusions following from their findings. Gade 2008, Higginson 2009, Higginson 2014, McCaffrey 2013 and Sahlen 2016 (linked to Brannstrom 2014) provided details of statistical tests and confidence intervals.

Consensus on Health Economic Criteria (CHEC) list

We also used the CHEC list to assess the methodological quality of economic evaluations (see Appendix 9). Overall, 13 studies met seven to 16 (out of 19) quality items on the list. Five items were considered to have been met by all studies: clear description of study population; a well‐defined research question in answerable form; identification of important and relevant outcomes for each alternative; appropriate measurement of outcomes; and conclusion following the reported data. All studies but Mendoza‐Galindo 2018 (abstract only) discussed the generalisation of results to other settings or patient group and chose the appropriate time horizon to include relevant costs and outcomes. Eleven of 13 studies used the appropriate economic study design to answer the stated objective with the exception of Brumley 2007 and Brannstrom 2014. All studies except McCaffrey 2013 and Mendoza‐Galindo 2018 (abstract only) discussed the ethical and distributional issues appropriately. Only two studies clearly described the competing alternatives (Higginson 2014; Ozcelik 2014), and three studies were considered to have appropriately chosen a perspective for the study (Higginson 2014; McCaffrey 2013; Temel 2010). Valuing outcomes appropriately was achieved only in five studies (Farquhar 2014; Farquhar 2016; Kane 1984; McCaffrey 2013; Temel 2010). No study needed nor clearly stated the discounting methods.

Patient Health‐related Quality of Life (HRQoL)

Ten studies contributed adjusted endpoint data to the main meta‐analysis on patient HRQoL (Bakitas 2009; Bakitas 2015; El‐Jawahri 2016; McCorkle 2015; O'Riordan 2019; Rodin 2019; Tattersall 2014; Temel 2010; Temel 2017; Vanbutsele 2018). We also pooled nine studies reporting unadjusted endpoint data (Bajwah 2015; Brannstrom 2014; El‐Jawahri 2016; Franciosi 2019; Gade 2008; Higginson 2014; Jingfen 2017; McCorkle 2015; Rogers 2017), and nine studies presenting unadjusted change data in our sensitivity analyses (Bajwah 2015; Bekelman 2018; El‐Jawahri 2016; Grudzen 2016; Ozcelik 2014; Rogers 2017; Sidebottom 2015; Temel 2010; Temel 2017). We further carried out sensitivity analyses to explore the effect of using 0.02 in adjusting for clustering in McCorkle 2015 among studies that reported adjusted endpoint data and unadjusted endpoint data. Only Solari 2018 reported adjusted change data.

Of the remaining 19 studies that were not in any of the meta‐analyses, 10 did not report on patient HRQoL (Ahronheim 2000; Brumley 2007; Carson 2016; Cheung 2010; Higginson 2009; Kane 1984; Ma 2019; McCaffrey 2013; Mendoza‐Galindo 2018 (abstract only); Wallen 2012), six presented data on different domains of HRQoL (Edmonds 2010; Farquhar 2014; Farquhar 2016; Groenvold 2017; Janssens 2019; Lowther 2015), one assessed HRQoL only at baseline but not at follow‐up (Hopp 2016), while one study only reported that there was "no significant difference" without presenting data (McWhinney 1994). Nottelmann 2018 assessed HRQoL but did not present analysable data.

Pooled data from 10 studies reporting adjusted endpoint data (main meta‐analysis) with 1344 participants showed that HSPC was beneficial at improving patient HRQoL when compared to usual care (SMD 0.26, 95% CI 0.15 to 0.37; I² = 3%; Analysis 1.1). Positive SMDs indicate better patient HRQoL while negative SMDs indicate lower patient HRQoL. The effect size obtained (0.26) is small based on conventional standards (Cohen 1988).

We carried out sensitivity analysis with studies that reported adjusted endpoint data to assess the impact of using an estimate of 0.02 in adjusting for clustering in McCorkle 2015. We found similar results to the main analysis, in favour of HSPC (SMD 0.29, 95% CI 0.18 to 0.40; I² = 0%; n = 9 studies; N = 1280 participants; Analysis 1.2). Sensitivity analysis using unadjusted endpoint values led to a larger difference between groups but the confidence intervals were wider and there was greater heterogeneity (SMD 0.41, 95% CI 0.11 to 0.70; I² = 83%; n = 9 studies; N = 1201 participants; Analysis 1.3). When McCorkle 2015 was removed, HSPC was still better than usual care in improving HRQoL (SMD 0.46, 95% CI 0.13 to 0.78; I² = 85%; n = 8 studies; N = 1137 participants; Analysis 1.4). When we pooled unadjusted change values, we also found benefit with HSPC (SMD 0.67, 95% CI 0.16 to 1.18; I² = 95%; n = 9 studies; N = 1278 participants; Analysis 1.5). The results from these sensitivity analyses supported that from the main analyses.

Solari 2018 was the only study that presented adjusted change data and it assessed patient HRQoL using the Schedule for the Evaluation of Individual Quality of Life ‐ Direct Weighting (SEIQoL‐DW) (range, 0 to 100, 100 = best HRQoL). It found no between‐group difference between HSPC and usual care both at three months and six months. At three months, mean change in the HSPC group was ‐0.9 (95% CI ‐6.8 to 5.1) and ‐3.7 (95% CI ‐17.6 to 10.3) in the usual care group with a difference of 2.8 (95% CI ‐12.2 to 17.8) between the groups. At six months, mean change in the HSPC group was 0.8 (95% CI ‐5.3 to 6.9) and that in the usual care group was ‐4.0 (95% CI ‐21.1 to 13.1) with a difference of 4.8 (95% CI ‐13.2 to 22.7) between the groups.

Across the studies in the meta‐analyses, we combined different scales assessing patient HRQoL by calculating SMDs. Table 3 under Additional tables describes the HRQoL scales and the dimensions they covered. The studies used different scales for measuring patient HRQoL (Kings Brief Interstitial Lung Disease in Bajwah 2015; Functional Assessment of Chronic Illness therapy for Palliative Care, FACIT‐Pal, in Bakitas 2009, Bakitas 2015 and Rogers 2017; Kansas City Cardiomyopathy Questionnaire (KCCQ) in Bekelman 2018; EQ‐5D in Brannstrom 2014; Functional Assessment of Cancer Therapy‐Bone Marrow Transplant, FACT‐BMT, in El‐Jawahri 2016; Functional Assessment of Chronic Illness Therapy ‐ Spiritual Well‐being Scale, FACIT‐Sp, in Rodin 2019; Modified City of Hope Patient Questionnaires in Gade 2008; Functional Assessment of Cancer Therapy‐General Measure, FACT‐G, in Franciosi 2019; Grudzen 2016 and McCorkle 2015; Chronic Respiratory Disease Questionnaire‐Health Related Quality of lIfe (CRQ HRQL) in Higginson 2014; European Organisation for the Research and Treatment of Cancer Quality of Life Questionnaire Core 30, EORTC QLQ‐C30 (Chinese version), in Jingfen 2017; EORTC QLQ‐C30 in Ozcelik 2014 and Vanbutsele 2018; Minnesota Living with Heart Failure Questionnaire (MLHF) in O'Riordan 2019 and Sidebottom 2015; Schedule for the Evaluation of Individual Quality of Life ‐ Direct Weighting (SEIQoL‐DW) in Solari 2018; McGill Quality of Life Questionnaire in Tattersall 2014; Functional Assessment of Cancer Therapy – Lung scale, FACT‐L, in Temel 2010; Functional Assessment of Cancer Therapy ‐ General scale, FACT‐G, in Temel 2017).

Four studies used more than one scale to measure patient HRQoL (Bajwah 2015; Brannstrom 2014; Higginson 2014; Rogers 2017). In particular, Brannstrom 2014 only showed data obtained using the EQ‐5D and not that from the KCCQ. Consequently, data from the EQ‐5D was used in the meta‐analysis. Higginson 2014 assessed HRQoL using the CRQ HRQL and the EQ‐5D. We only used data from the CRQ HRQL in the meta‐analysis because unlike the EQ‐5D (Williams 1995), a generic health‐related quality of life measure, it is more specific to chronic respiratory disease (Guyatt 1987). Rogers 2017 assessed HRQoL using the FACIT‐Pal and the KCCQ and both were presented as primary outcomes. Given that the FACIT‐Pal has more extensive validation in palliative populations, we used it in the meta‐analysis.

Overall, the funnel plot suggested some asymmetry (Figure 7). Egger's test for asymmetry resulted in a P value of 0.02. However, given evidence of publication of negative studies in the funnel plot, this asymmetry is not necessarily indicative of publication bias. We did not carry out subgroup analysis due to low heterogeneity (I² = 3%) in our main meta‐analysis.

Figure 7

---

---

Funnel plot of comparison: 1 Patient health‐related quality of life, outcome: 1.1 HSPC versus usual care on patient HRQoL: adjusted endpoint values.

Patient symptom burden (as a collection of two or more symptoms)

We pooled six studies that reported adjusted endpoint data as the main meta‐analysis on patient symptom burden (Bakitas 2009; Bakitas 2015; El‐Jawahri 2016; Rodin 2019; Tattersall 2014; Temel 2010). We pooled six studies that reported unadjusted endpoint values (Bajwah 2015; El‐Jawahri 2016; Gade 2008; Higginson 2014; Lowther 2015; McCorkle 2015), four studies presenting adjusted change values (Edmonds 2010; McCorkle 2015; Sidebottom 2015; Solari 2018), and six studies that reported unadjusted change values in our sensitivity analyses (Bajwah 2015; Bekelman 2018; El‐Jawahri 2016; Higginson 2009; Ozcelik 2014; Temel 2010). We further carried out sensitivity analyses to explore the effect of using 0.02 in adjusting for clustering in McCorkle 2015 among studies that reported unadjusted endpoint data and adjusted change data.

Pooled data from six studies with 761 participants reporting adjusted endpoint values showed that HSPC was beneficial at reducing patient symptom burden compared to usual care (SMD ‐0.26, 95% CI ‐0.41 to ‐0.12; I² = 0%; Analysis 2.1). Negative SMDs indicate benefit (lower symptom burden) and positive SMDs reflect higher symptom burden.

Sensitivity analysis in the six studies (N = 833 participants) that reported unadjusted endpoint values showed a pooled effect of SMD ‐0.17 (95% CI ‐0.54 to 0.20; I² = 83%; Analysis 2.2). Among studies that reported unadjusted endpoint values, we carried out another sensitivity analysis to assess the impact of using 0.02 in adjusting for clustering in McCorkle 2015 and had similar findings (SMD ‐0.19, 95% CI ‐0.62 to 0.24; I² = 87%; n = 5 studies; N = 769 participants; Analysis 2.3). When we considered adjusted change values, the pooled effect was a SMD of ‐1.31 (95% CI ‐3.27 to 0.64; I² = 98%; n = 4 studies; N = 353 participants; Analysis 2.4). When we excluded McCorkle 2015 from the studies that reported adjusted change values, we found a pooled effect of SMD ‐1.79 (95% CI ‐4.29 to 0.70; I² = 98%; n = 3 studies; N = 289 participants; Analysis 2.5). When we considered unadjusted change values, the pooled effect from the studies was a SMD of ‐0.44 (95% CI ‐0.94 to 0.06; I² = 88%; n = 6 studies; N = 641 participants; Analysis 2.6).

Of the remaining 25 studies that were not in any of the meta‐analyses, 20 did not report on patient symptom burden (Ahronheim 2000; Brumley 2007; Carson 2016; Cheung 2010; Franciosi 2019; Farquhar 2014; Farquhar 2016; Groenvold 2017; Grudzen 2016; Hopp 2016; Jingfen 2017; Ma 2019; McCaffrey 2013; McWhinney 1994; Mendoza‐Galindo 2018 (abstract only); Nottelmann 2018; Rogers 2017; Temel 2017; Vanbutsele 2018; Woo 2019), two studies reported that there were "no significant differences" between intervention and control groups but they did not present data (Brannstrom 2014; Kane 1984), while O'Riordan 2019 did not present data from the Edmonton Symptom Assessment Scale (ESAS). Wallen 2012 did not present analysable data, while Janssens 2019 only assessed patient symptom burden using the Edmonton Symptom Assessment Scale (ESAS) in the intervention group.

Across the studies that we pooled in the meta‐analyses, we combined different generalised measures of patient symptom burden by applying SMDs. Included studies used the following measures in assessing patient symptom burden: Palliative care Outcome Scale (POS) or a modified form of it in Bajwah 2015, Edmonds 2010, Higginson 2009, Higginson 2014 and Solari 2018; African POS in Lowther 2015; Edmonton Symptom Assessment Scale (ESAS) or a modified form of it in Bakitas 2009, El‐Jawahri 2016, Ozcelik 2014 and Sidebottom 2015; symptom impact subscale of the Quality of Life at End of life (QUAL‐E) in Bakitas 2015; General Symptom Distress Scale in Bekelman 2018; physical area scale of the Modified City of Hope Patient Questionnaires (MCOHPQ) in Gade 2008; Symptom Distress Scale (SDS) in McCorkle 2015 and Wallen 2012; Rotterdam Symptom Checklist (RSC ‐ Physical Symptoms Score) in Tattersall 2014; lung cancer subscale (LCS) of the FACT‐L in Temel 2010 and Memorial Symptom Assessment Scale (MSAS) in Rodin 2019. Only the severity subscale of the MSAS reported by Rodin 2019 was used in the meta‐analysis.

Given that there were fewer than 10 included studies in the main meta‐analysis of studies that presented adjusted endpoint values, we did not use funnel plots or carry out tests for funnel plot asymmetry. We also did not carry out subgroup analysis due to lack of heterogeneity I² = 0%) in our main meta‐analysis.

Patient satisfaction with care

Eight studies assessed the effect of HSPC on patient satisfaction with care (Brumley 2007; Gade 2008; Jingfen 2017; Kane 1984; O'Riordan 2019; Ozcelik 2014; Rodin 2019; Wallen 2012). We excluded three of the studies from the synthesis because they used measures that had not been validated (Jingfen 2017; O'Riordan 2019; Ozcelik 2014), while one study did not present analysable data (Wallen 2012). We excluded Janssens 2019 because the authors did not state the outcome measure used in assessing satisfaction with the intervention. The remaining four studies with 733 participants used validated measures (Brumley 2007; Gade 2008; Kane 1984; Rodin 2019). However, we could not include Brumley 2007 and Kane 1984 in our meta‐analysis because Brumley 2007 presented odds ratio while Kane 1984 only presented P values.

Gade 2008 and Rodin 2019 reported adjusted endpoint values and found evidence in favour of HSPC (SMD 0.36, 95% CI 0.14 to 0.57; I² = 0%; N = 337 participants; Analysis 3.1). Positive SMDs indicate better patient satisfaction while negative SMDs indicate lower patient satisfaction. Gade 2008 used the Modified City of Hope Patient Questionnaires (MCOHPQ) Place of Care Environment Scale and the Doctors, Nurses/Other Care Providers Communication scale for assessing patient satisfaction with care. The MCOHPQ Place of Care Environment scale addressed experiences receiving pain management and symptom relief, psychological and social support, discharge planning, and end‐of‐life planning, while the Doctors, Nurses/Other Health Care Providers Communication scale addressed the level of caring and respect a patient felt from their providers, as well as the opportunity, ease, and the level of understanding the patient had with their providers. Only data from the MCOHPQ Place of Care Environment scale was used in the meta‐analysis. Rodin 2019 assessed patient satisfaction with care using the 16‐item Family Satisfaction with Care ‐ Patient Version (FAMCARE‐P16).

Brumley 2007 found a 3.37 higher odds of satisfaction in the HSPC group compared to control group (P = 0.03). Brumley 2007 used the Reid‐Gundlach Satisfaction with Services instrument for assessing patient satisfaction. Kane 1984 found differences in satisfaction scores (P < 0.01) with HSPC patients expressing more satisfaction than control patients in two of the three areas examined. The two areas were interpersonal care and involvement in care. Kane 1984 used the Interpersonal Care scale adapted from the Ware scale (Ware 1979), a physical environment scale from McCaffree and Harkins (McCaffree 1976) and involvement‐in‐care questions adapted from the National Cancer Institute’s Hospice Study (Baker 1981). Kane 1984 reported that these measures have been shown to be reliable and valid for patients with terminal cancer. No study reported decreased satisfaction of care in the HSPC group.

Due to small numbers in our main meta‐analysis with adjusted endpoint values, we could not carry out subgroup analysis and we did not use funnel plots or carry out tests for funnel plot asymmetry.

Unpaid caregiver satisfaction with care

Four studies assessed the effect of HSPC on family satisfaction with care (Carson 2016; Cheung 2010; Kane 1984; Ozcelik 2014). We excluded Cheung 2010 and Ozcelik 2014 from the synthesis because they used non‐validated family satisfaction measures. Carson 2016 and Kane 1984 used validated measures with a total of 408 participants.

Carson 2016 was the only study that presented adjusted endpoint values, with family satisfaction assessed using the Family Satisfaction in the Intensive Care Unit (FS‐ICU) survey (range, 0 to 100, 100 = best unpaid caregiver satisfaction). It found no between‐group difference between HSPC and usual care. The mean (95% CI) satisfaction in the HSPC group was 81.1 (78.3 to 83.9) while that in the usual care group was 84.3 (81.3 to 87.3), with a difference of ‐3.1 (‐7.3 to 1.0) between groups (P = 0.13).

Kane 1984 did not present their data. They only reported P values in favour of the HSPC group in two of the five cohorts they assessed. Kane 1984 assessed family satisfaction with care using the Interpersonal Care scale adapted from the Ware scale (Ware 1979), a physical environment scale based on that of McCaffree and Harkins (McCaffree 1976), and involvement‐in‐care questions adapted from the National Cancer Institute’s Hospice Study (Baker 1981).

Achieving patient preferred place of death (measured by number of patients with home death)

Given that most people in developed countries prefer to die at home (Gomes 2012), we used number of home deaths as a proxy measure for achieving patient preferred place of death.

Pooled data from seven studies with 861 analysed participants showed that those receiving HSPC had higher odds of home deaths compared to those receiving usual care (OR 1.63, 95% CI 1.23 to 2.16; I² = 0%; Analysis 4.1). The odds ratio of 1.63 translates to a risk ratio of 1.22 (95% CI 1.08 to 1.39). This implies an increase in the relative risk of home deaths of 22% (95% CI 8% to 39%) when compared to usual care.

Kane 1984 reported that in the intervention group, only 3% of deaths occurred at home with almost 60% dying in the inpatient hospice, while in the control group, 7% of deaths occurred at home with almost 80% dying in hospital. The actual number of deaths was not given but the authors stated that the difference between intervention and control group was not "statistically significant". Janssens 2019 reported two home deaths but did not state whether they occurred in the HSPC group or control group.

The remaining 33 studies did not report on home death.

Given that there were fewer than 10 included studies in the meta‐analysis, we did not use funnel plots or carry out tests for funnel plot asymmetry. In addition, we could not carry out subgroup analysis due to lack of heterogeneity (I² = 0%) in our meta‐analysis.

Achieving patient preferred place of care

Only one study by Bajwah 2015 (n = 47 participants) reported on this outcome. Bajwah 2015 was a fast‐track RCT. Patients in the intervention group received HSPC immediately after randomisation, while the control group received HSPC four weeks after randomisation. Consequently, both the intervention and control group received HSPC. Results at the end of the study showed that all eight patients (100%) who died in the intervention group achieved their preferred place of care, while 11 patients (84%) in the control group who received HSPC after four weeks achieved this.

Mortality/survival

Thirty‐six studies with 7103 participants reported on mortality/survival (Ahronheim 2000; Bajwah 2015; Bakitas 2009; Bakitas 2015; Bekelman 2018; Brannstrom 2014; Brumley 2007; Carson 2016; Cheung 2010; Edmonds 2010; El‐Jawahri 2016; Farquhar 2014; Farquhar 2016; Franciosi 2019; Gade 2008; Groenvold 2017; Grudzen 2016; Higginson 2009; Higginson 2014; Hopp 2016; Janssens 2019; Kane 1984; Lowther 2015; Ma 2019; McCaffrey 2013; McCorkle 2015; McWhinney 1994; O'Riordan 2019; Rogers 2017; Sidebottom 2015; Solari 2018; Tattersall 2014; Temel 2010; Temel 2017; Vanbutsele 2018; Woo 2019) (see Table 4 under Additional tables). We decided against pooling of their hazard ratios in a meta‐analysis due to methodological limitations in the included studies. Three studies did not report on the number of deaths (Mendoza‐Galindo 2018 (abstract only); Ozcelik 2014; Wallen 2012), while Nottelmann 2018 only reported number of deaths in the HSPC group. Rodin 2019 reported that there were no deaths during the study, while this was unclear in the foreign language study because it was not described (Jingfen 2017).

Ten of these studies reported on deaths in the HSPC and control group without presenting survival time and they found no between‐group difference in number of deaths (Ahronheim 2000; Bekelman 2018; Brannstrom 2014; Cheung 2010; Franciosi 2019; Higginson 2009; Hopp 2016; Ma 2019; McCaffrey 2013; Rogers 2017), while Sidebottom 2015 reported no association between study group assignment and death within six months after adjustment for age, gender, and marital status (Hazard Ratio: 1.90 (95% CI: 0.88, 4.09); P = 0.101). Sidebottom 2015 reported 14 deaths (12.1%) in the HSPC group and 5 deaths (4.3%) in the control group.

In 11 studies, it was unclear if there was any difference in mortality because the P values were not presented (Bajwah 2015; Edmonds 2010; El‐Jawahri 2016; Farquhar 2014; Farquhar 2016; Lowther 2015; McCorkle 2015; McWhinney 1994; O'Riordan 2019; Solari 2018; Temel 2017). McWhinney 1994 only presented the total number of deaths at one month (36 (24.7%)) but did not report the numbers in the HSPC and control group.

In the studies that reported survival time, there was probably little to no effect of HSPC on survival (Bakitas 2009; Bakitas 2015; Carson 2016; Gade 2008; Groenvold 2017; Grudzen 2016; Kane 1984; Janssens 2019; Vanbutsele 2018; Woo 2019).

In Bakitas 2009, median survival (95% CI) in the HSPC group was 14 months (10.6 to 18.4) and 8.5 months (7 to 11.1) in control with a P value of 0.14. There were 112 deaths (69.6%) in the HSPC group and 119 deaths (73.9%) in the control group. The Cox proportional hazards model estimate demonstrated a reduced relative risk of death (Hazard Ratio (HR): 0.67 (95% CI: 0.496 to 0.906), P = 0.009) in the HSPC group during the first year of the study and a greater relative risk after one year (HR, 1.56 (95% CI: 0.908 to 2.655)).

In Bakitas 2015, a fast‐track RCT in which the intervention group was offered HSPC immediately, while the control group received HSPC after three months, median survival by the end of data collection in the intervention group was 18.3 months and 11.8 months in the control group who began HSPC three months later. Kaplan‐Meier curves illustrated a 15% difference in survival at 1 year (HSPC, 63% versus control, 48%; P = 0.038). However, the overall log‐rank test P value was 0.18, suggesting a convergence in overall survival after 12 months. At one year, there were 109 deaths (52.7%) but numbers in intervention and control groups were not reported.

Carson 2016 reported a median survival (95% CI) of 19 (12 to 37) days in the HSPC group and 23 (12 to 39) days in control group (P = 0.51). There was no difference in 90‐day survival (HR, 0.95 (95% CI: 0.65 to 1.38), P = 0.96). Post hoc adjustment for baseline activities of daily living and study site did not alter the outcome (HR,1.01 (95% CI: 0.69 to 1.47), P = 0.96).

In Grudzen 2016, median survival (95% CI) in the HSPC group was 289 days (128 to 453) and 132 days (80 to 302) in control with a P value of 0.2. At one year, 41 participants (59.4%) had died in the HSPC group and 44 (65.7%) had died in the control group. However, there was no difference between the groups (P = 0.20).

Janssens 2019 was not clear about whether they were reporting mean or median survival. Survival in the HSPC group was 454 days (95% CI: 382 to 525) and 425 days (95% CI: 339 to 509) in the control group (log‐rank test, P = 0.91). During the follow‐up period in Janssens 2019, there were four deaths (15.4%) in the HSPC group and four deaths (17.4%) in the control group.

Kane 1984 reported no difference in survival between HSPC and the control group as the survival curves were similar.

In Gade 2008, median survival (IQR) was 30 days (6 to 104) in the HSPC group and 36 days (13 to 106) in the control group (P = 0.08). There were 173 deaths (63%) in the HSPC group and 132 deaths (56%) in the control group during the study period.

Groenvold 2017 reported that survival time did not differ between HSPC and the control group. Median survival in the HSPC group was 323 days and 364 days in the control group (P = 0.16, but in the adjusted analysis P = 0.39). There were 25 deaths (27%) in the HSPC group and 22 deaths (23%) in the control group.

Woo 2019 reported that there was no difference in survival between HSPC and usual care but did not present any data.

Vanbutsele 2018 found the median survival (95% CI) in the HSPC group to be 312 days (190 to 434) and 343 days (253 to 433) in the control group (P = 0.97).

Sidebottom 2015 reported no association between study group assignment and death within six months after adjusting for age, gender and marital status (P = 0.10).

Higginson 2014 and Temel 2010 found evidence in favour of HSPC for longer survival compared to usual care. Higginson 2014 was a fast‐track RCT in which the intervention group received HSPC immediately while those in the control group were offered HSPC after six weeks. Survival was calculated from the time of randomisation to the time of death, if death occurred during the study period, or to the time of censoring. Median survival (range) from randomisation to the time of censoring was 745 (338 to 1075) days in the intervention group compared to 711 (345 to 1045) in the control group who received HSPC after six weeks (P = 0.048). In subgroup analysis, this pattern was not recorded for patients with cancer (P = 0.97); but it became more marked for patients with diseases other than cancer (P = 0.01). Temel 2010 reported that median survival (95% CI) was 11.6 months (6.4 to 16.9) in the HSPC group and 8.9 months (6.3 to 11.4) in control (log rank P = 0.02). After adjustment for age, sex, and baseline Eastern Cooperative Oncology Group performance status, the group assignment remained a predictor of survival (hazard ratio for death in the standard care group, 1.70; 95% CI, 1.14 to 2.54; P = 0.01).

By contrast, Brumley 2007 and Tattersall 2014 reported greater survival (SD) in the control group compared to the HSPC group. Brumley 2007 reported a mean (SD) survival of 242 (SD:200) days in the control group compared to 196 (SD:164) days in those receiving HSPC (P = 0.03). However, results of the Kaplan‐Meier survival analysis did not show differences in survival time between study groups (P = 0.08). The authors also highlighted 75% death among participants but the percentages in the HSPC and control groups were not stated. In Tattersall 2014, there were 39 (65%) deaths in the HSPC group and 31 (51.7%) in the control group at 12 months. Tattersall 2014 found the median survival (95% CI) in the HSPC group to be 7 months (5.2 to 9.8) compared to 11.7 months (9.8 to 18.8) in the control group (log rank P = 0.014). The estimated hazard ratio was 1.6 (95% CI:1.1 to 2.3; P = 0.015). This estimate changed to 1.5 (95% CI 0.99 to 2.2; P = 0.06) when adjusted for the oncologist’s baseline estimate of likely survival, diagnosis, months since diagnosis, and gender.

Pain

We pooled data from four studies (n = 525 participants) that reported adjusted endpoint values for pain as the main meta‐analysis. The meta‐analysis showed that HSPC may lead to little to no difference in pain relief (SMD ‐0.16, 95% CI ‐0.33 to 0.01; I² = 0%; Analysis 5.1). Positive SMDs indicate more pain while negative SMDs indicate lower pain (benefit). Only Woo 2019 reported unadjusted endpoint values and it assessed pain using the Brief Pain Inventory. It found no difference in mean pain scores between HSPC and usual care (P = 0.22). However, sensitivity analysis with studies reporting adjusted change values showed evidence in favour of HSPC (SMD ‐0.47, 95% CI ‐0.74 to ‐0.20, I² = 0%; n = 2 studies; N = 218 participants; Analysis 5.2).

When we carried out sensitivity analysis using unadjusted change values, we found no evidence of a difference between HSPC and usual care (SMD ‐0.93, 95% CI ‐3.05 to 1.19; I² = 97%; n = 2 studies; N = 291 participants; Analysis 5.3).

Although we had initially specified that we would treat pain as a binary outcome in our published protocol (Bajwah 2017), this was not possible as most studies presented pain as a continuous outcome. Studies such as Tattersall 2014 reported on the percentage of patients with pain, while Lowther 2015 presented pain data as medians. Kane 1984 reported that there was no difference in pain between the intervention and control group over time but did not present data. Also, McWhinney 1994 stated that there were "no clinically or statistically significant differences" between the intervention and control groups but did not report their data. The remaining 30 studies did not report on pain.

We combined different scales assessing pain by calculating SMDs. Across the studies in these meta‐analyses, we combined different measures for assessing pain (PEG derived from the Brief Pain Inventory (BPI) in Bekelman 2018; pain item of the EORTC QLQ‐C30 in Groenvold 2017 and Vanbutsele 2018; pain item of the POS in Higginson 2009; pain severity on the BPI in O'Riordan 2019, Rodin 2019 and Woo 2019; pain item of the ESAS in Ozcelik 2014 and Sidebottom 2015).

Given that there were fewer than 10 included studies in our main meta‐analysis on pain using adjusted endpoint values, we did not use funnel plots or carry out tests for funnel plot asymmetry. In addition, we could not carry out subgroup analysis due to lack of heterogeneity (I² = 0%) in our main meta‐analysis with adjusted endpoint values.

Patient anxiety and depression

Patient breathlessness

We pooled data from five studies reporting adjusted endpoint values for our main meta‐analysis on patient breathlessness with a pooled estimate of SMD ‐0.04 (95% CI ‐0.19 to 0.12; I² = 0%, N = 616 participants; Analysis 12.1). Negative SMDs indicate benefit (reduced breathlessness) and positive SMDs reflect worsened breathlessness. The five studies used different instruments and reported on different breathlessness domains. For instance, Farquhar 2014 and Farquhar 2016 both assessed distress due to breathlessness and breathlessness mastery using a Numeric Rating Scale (NRS) and the mastery domain of the Chronic Respiratory Questionnaire (CRQ), respectively; Groenvold 2017 and Vanbutsele 2018 assessed breathlessness intensity using the dyspnoea item of EORTC QLQ‐C30; O'Riordan 2019 assessed breathlessness intensity using the BORG scale. For Farquhar 2014 and Farquhar 2016, we used only data for distress due to breathlessness assessed with the NRS in our meta‐analysis because it was the primary outcome. We did not differentiate between different breathlessness domains in our meta‐analysis due to small numbers.

Sensitivity analysis carried out with the two studies (N = 128 participants) presenting unadjusted endpoint values showed a pooled estimate in favour of HSPC (SMD ‐0.35, 95% CI ‐0.70 to ‐0.00; I² = 0%; Analysis 12.2).

Only Sidebottom 2015 presented adjusted change values. It assessed breathlessness using the dyspnoea item of ESAS (using a visual scale line, 0 to 10, 10 = worst possible) and found that breathlessness scores improved by a mean of 2.8 points in the HSPC group and 1.7 in the control group at 3 months (difference 1.08, P < 0.001) after adjusting for age, gender, and marital status differences between study groups. This difference was evident at one month with a mean difference of 1.10 (P < 0.001).

Sensitivity analysis with the two studies that reported unadjusted change values showed a pooled estimate of SMD ‐0.47 (95% CI ‐1.55 to 0.61; I² = 90%, N = 292 participants; Analysis 12.3).

A study by Tattersall 2014 also recorded this outcome but did not present analysable data. The remaining 31 studies did not report on breathlessness.

Studies included in the meta‐analyses used different scales in assessing breathlessness: D‐12 in Bajwah 2015; Memorial Symptom Assessment Scale in Bekelman 2018; Numeric Rating Scale (NRS) for distress due to breathlessness in Farquhar 2014 and Farquhar 2016; dyspnoea item of EORTC QLQ‐C30 in Groenvold 2017 and Vanbutsele 2018; breathlessness mastery domain of the Chronic Respiratory Disease Questionnaire (CRQ mastery) in Higginson 2014; BORG scale in O'Riordan 2019; dyspnoea item of ESAS in Ozcelik 2014 and Sidebottom 2015.

Due to lack of heterogeneity (I² = 0%) in our main meta‐analysis, we could not carry out subgroup analysis. Given that there were fewer than 10 included studies in the main meta‐analysis on breathlessness using adjusted endpoint values, we did not use funnel plots or carry out tests for funnel plot asymmetry.

Adverse events in patients and unpaid caregivers

Eight studies with 1252 participants reported on adverse events (Bajwah 2015; Bekelman 2018; Groenvold 2017; Higginson 2014; Lowther 2015; Rodin 2019; Solari 2018; Tattersall 2014) (see Table 5 under Additional tables). Two of these studies involved unpaid caregivers (Bajwah 2015; Higginson 2014).

Six studies (N = 976 participants) reported no harmful effect (Bajwah 2015; Bekelman 2018; Groenvold 2017; Higginson 2014; Lowther 2015; Rodin 2019).

One study by Tattersall 2014 (N = 120 participants) found that more patients in the HSPC group had the mild adverse event of poorer appetite (P = 0.04) compared to the control group.

Solari 2018 (N = 156 participants) reported 15 serious adverse events in 13 patients in the HSPC group and seven in seven patients in the control group (P = 0.78). Serious adverse events reported included aspiration pneumonia, generalised anxiety, breathing difficulty, urine retention/infection, anarthria, contact dermatitis, dysphagia, vomiting, bladder catheter malfunctioning, fever, arrhythmia, necrotising fasciitis, traumatic wound, macrohaematuria, constipation, abdominalgia and bronchitis. Three patients in the HSPC group died but this was considered to be unrelated to the intervention.

Unpaid caregiver symptom control

Unpaid caregiver pre‐ and post‐bereavement outcomes

Resource use

It was not possible to combine data for resource use or costs due to differences in measurement and reporting, such as type of analysis, tools used, assessment time points or time horizon and statistics reported. Consequently, we provided a narrative synthesis on the economic studies.

Thirty‐one studies compared resource use or costs or both between the treatment groups in different ways. Three studies collected information on resource use and/or costs by chart review (Ahronheim 2000; Kane 1984; Bakitas 2009), while four studies collected resource use data from patients using either the Client Services Receipt Inventory (CSRI) or a modified form of it (Farquhar 2014; Farquhar 2016; Higginson 2009; Higginson 2014). Eight studies used medical/health records (Grudzen 2016; Ma 2019; Rogers 2017; Sidebottom 2015; Tattersall 2014; Temel 2010; O'Riordan 2019; Vanbutsele 2018). Four studies used a combination of methods (Bekelman 2018; Bakitas 2015; Janssens 2019; Rodin 2019). Bekelman 2018 collected data from medical records and supplemented these with patient or family self‐report, while Janssens 2019 collected data from medical records as well as contact with patients and their GPs. Rodin 2019 collected data from patients and their medical charts. Bakitas 2015 used patient self‐report for hospital and intensive care unit days and emergency department visits, while decedents' data for the period between the last patient‐reported assessment and death, and chemotherapy use in last 14 days were obtained from medical records. In Ozcelik 2014, a patient expenditure record form was created to capture resources and their costs. Brumley 2007 obtained resource use for each patient retrospectively from the non‐profit HMO mainframe database, while Gade 2008 used standard data extract protocols to extract information from the managed care organisation’s (MCO) database. Methods for collecting resource use information were unclear in nine RCTs (Brannstrom 2014; Cheung 2010; Carson 2016; El‐Jawahri 2016; Groenvold 2017; McCaffrey 2013; Mendoza‐Galindo 2018 (abstract only); Temel 2017; Woo 2019).

We considered resource use in the following areas: institutional care services use, outpatient clinic services use, community care services use, unpaid caregiver's care, and medications and other resources.

Costs and cost‐effectiveness of HSPC

Thirteen economic studies with 2103 participants reported on cost. The utilisations included were: ED or A&E visits; inpatient and outpatient hospital care; home and community care; care in nursing homes (or skilled nursing homes); inpatient stay; day care in hospice; hospice care at home; informal care; drugs and equipment. Four studies reported the results of cost‐effectiveness analysis using relevant outcome measures (palliative outcome, unpaid caregiver’s burden, quality‐adjusted‐life‐years) and hospital costs or total costs (Farquhar 2014; Farquhar 2016; Higginson 2009; McCaffrey 2013). Results of cost‐effectiveness analyses were reported by ICERs and/or costs per QALY (point estimates or cost‐effectiveness planes).

Two studies found evidence of lowered cost with HSPC (Brumley 2007; Gade 2008). When compared to usual care, Mendoza‐Galindo 2018 (abstract only) reported a reduction in the cost of hospitalisation days in the HSPC group. However, no difference was found between groups in the cost of emergency room visits. In Brannstrom 2014, this was unclear as no P value was presented for the difference in cost between HSPC and usual care. We identified four full economic studies (Farquhar 2014; Farquhar 2016; Higginson 2009; McCaffrey 2013). The evidence on the cost‐effectiveness of HSPC compared to usual care was inconsistent.

With the exception of Mendoza‐Galindo 2018 (abstract only), all other studies had applied more robust methodology since the first relevant study we identified which compared the costs of HSPC and conventional care among cancer patients (Kane 1984). Kane 1984 provided services across multiple settings and was carried out in the USA. The HSPC group had lower total costs when compared to conventional care. However, the authors reported that the difference was "not significant". The estimated mean expenditure per patient was reported to be US dollars (USD) 15,263 (converts to Great British Pounds (GBP) 29,058 in 2018) in the HSPC group and USD 15,493 (converts to GBP 29,496 in 2018) in the conventional care group. Resource use was measured in hospital stays, hospice stays, surgical procedures, chemotherapy and radiotherapy, and costs were calculated using different assumptions. However, difference in survival (days since enrollment in the study) as well as other factors (e.g. age, severity of diseases) which might be associated with costs, were not adjusted for.

Brumley 2007 compared resource use and costs between the HSPC and usual care group versus usual care only among terminally ill patients with mixed cancer and non‐cancer diagnoses in the USA involved service provision across multiple settings. A wider range of resource use was reported from the health insurance database: the number of ED visits, physician office visits, hospital days, hospice days, skilled nursing facility days, home health and palliative visits and palliative physician home visits. Service utilisation was lower in the HSPC group than usual care group even after controlling for age, survival and severity measured using the Palliative Performance Scale. Stay in hospital decreased by 4.36 days and ED visits by 0.35. Due to the difference in the survival (days on service), mean costs per patient were adjusted using regression analysis, controlling for survival, age, severity of illness and primary disease. Mean costs per patient in the intervention group were much lower (Australian dollars (AUD) 12,670, SD AUD 12,523; converts to GBP 8383, SD GBP 8285 in 2018), compared with the usual care group (AUD 20,222, SD AUD 30,026; converts to GBP 13,379, SD GBP 19,866 in 2018). Average daily costs per patient were also lower in the intervention group (AUD 95.30, converts to GBP 63.05 in 2018) compared to the usual care group (AUD 212.80, converts to GBP 140.76 in 2018) (P = 0.02).

Gade 2008 used the health insurance database to extract resource use and unit cost of services of hospitalised patients with life‐limiting illnesses (mixed cancer and non‐cancer diagnoses), who were randomly assigned to the HSPC intervention or usual care. Included utilisations were ED visits, clinic and hospital outpatient visits, home health visits, hospital admission, skilled nursing facility admissions and prescriptions. The cost of the palliative care team was calculated as the intervention cost. HSPC patients stayed longer in hospice after the index hospitalisation (24 days) than usual care patients (12 days) however this was a non‐significant difference (P = 0.08), and had shorter ICU stays on readmission (12 times versus 21 times, P = 0.04) and lower total healthcare costs (USD 14,486, converts to GBP 15,013 in 2018 versus USD 21,252, converts to GBP 22,025 in 2018, P = 0.001). Gade 2008 involved an inpatient consult model and was a USA study.

Temel 2010 examined the effectiveness of early palliative care integrated with standard oncologic care among patients with newly diagnosed metastatic non–small‐cell lung cancer, where standard oncologic care alone was a comparator. It was an outpatient model of HSPC that took place in the USA. Data on health utilisations and end‐of‐life care were collected from the medical records: anticancer therapy, medication prescriptions, referral to hospice, hospital admissions and ED visits. Patients in standard care received more aggressive end‐of‐life care (54% [30 of 56 patients] versus 33% [16 of 49 patients], P = 0.05), and had non‐significant longer stays in hospice care (median 11 days versus 4 days, P = 0.09) than the intervention group. Patients in the HSPC group used more palliative care and less aggressive care while there was greater improvement in quality of life and survival in this group than control. However, this was not conclusive because the sample size of the study did not allow the statistical power to test the differences in service utilisation. Detailed analyses of costs and cost‐effectiveness were conducted and reported later although lacking in statistical power to detect the difference in Greer 2016 (linked to Temel 2010)). Comparisons of costs per day alive and costs for the last 30 days were made between HSPC and the usual care group and the cost‐effectiveness per life year saved was calculated. Total costs per day were on average lower in the HSPC group. However, this was a non‐significant difference (mean difference USD 117, SE USD 74; converts to GBP 103, SE GBP 65 in 2018, P = 0.13). Total costs for the last 30 days were also reduced non‐significantly (mean difference USD 2527, SE USD 3311; converts to GBP 2230, SE GBP 2922 in 2018, P = 0.44). The cost‐effectiveness ratio was USD 41,938 per life year saved. There was a non‐signifant increase in use of hospice care (mean difference USD ‐1053, SE USD 538; converts to GBP ‐929, SE GBP 475 in 2018, P = 0.07) and less use of chemotherapy (mean difference USD 757, SE USD 365; converts to GBP 668, SE GBP 322 in 2018, P = 0.03) for the last 30 days.

Higginson 2014 examined the effectiveness of the early introduction of palliative care among patients with chronic breathlessness in the UK. The intervention (HSPC) was provided across multiple settings and patients had mixed cancer and non‐cancer diagnoses. Patients were randomly assigned either to the HSPC group or to usual care. Resource use was collected using the Client Services Receipt Inventory (CSRI) on health, voluntary, and social care received over the past three months at baseline and since the last interview at six weeks follow‐up. Limited results on resource use and costs were reported: hospital inpatient stays (mean 4.5, SD 6.8 in BSS; mean 4.6, SD 7.6 in control) and costs of formal care use (mean GBP 1422, 95% CI: 897 to 2101; converts to GBP 1611, 95% CI: 1016 to 2380 in 2018 in Breathlessness Support Service (BSS) group; mean GBP 1408, 95% CI: 899 to 2023; converts to GBP 1595, 95% CI: 1018 to 2292 in 2018 in control group). There was no between‐group difference between the two groups.

Brannstrom 2014 compared service use between patients randomised to the integrated palliative advanced home care and heart failure care (PREFER) intervention and usual care among patients with severe chronic heart failure in Sweden. The PREFER intervention involved an outreach model of HSPC. Resource use collected included hospital admissions, inpatient days, physician and nurse visits, phone calls and drug prescriptions. The HSPC group had fewer hospitalisations than the control group (0.42 ± 0.60 versus 1.47 ± 1.81, P = 0.009) and the length of stay in hospital was also shorter in patients receiving the intervention (mean 2.9, SD 8.3 versus mean 8.5, SD 12.4, P = 0.011). The total days or total contacts per study arm were compared between HSPC and the control group and additional cost analysis was reported in Sahlen 2016 (linked to Brannstrom 2014). QALY gain was 0.25 years between baseline and the end of intervention across the HSPC group (P = 0.025). Over six months, total cost was Swedish Krona (SEK) 1.4 million (EUR 140,000, converts to GBP 126,132 in 2018) in the HSPC group (n = 36) and SEK 2.0 million (EUR 205,000, converts to GBP 180,188 in 2018) in the control group (n = 26), and the difference SEK 600,000 (EUR 61,000, converts to GBP 54,056 in 2018) was the saving achieved by providing the intervention in addition to usual heart failure care.

Ozcelik 2014 compared duration of hospitalisation and direct cost between HSPC and usual care in Turkey. It was an inpatient consult model of HSPC. A patient cost record form was used to document cost. This form was created by listing direct health expenditure, which consisted of all expenses incurred while in hospital. Direct expenses assessed included medicines used from the start of the patient’s stay in hospital, medical equipment, laboratory and diagnosis tests, consultations, professional care, and hospital stay expenses (including those of companions). On the patient’s discharge from hospital, costs were recorded on the form by obtaining the expenses list from the clinic secretary. Mean (SD) direct cost in the HSPC group was USD 68,869 (SD 48.522) (converts to GBP 60,154 (GBP 42,382) in 2018) and USD 81,076 (72,700) (converts to GBP 70,816 (GBP 63,500) in 2018) in the control group (P = 0.76). There was no evidence of a difference in the duration of hospitalisation (P = 0.07), with a mean (SD) length of stay in hospital of 9.4 (6.27) days in the HSPC group and 13.9 (11.5) days in the control group.

The first study using robust cost‐effectiveness analysis (CEA) method among papers we identified was by Higginson 2009. This was the CEA alongside a feasibility trial of a new HSPC service among patients with multiple sclerosis (MS) in the UK, randomised into either fast‐track of the new intervention or control of usual care. Higginson 2009 involved service provision across multiple settings. Costs were measured in health, social and voluntary services, and informal care provided by family or friends was also included for the analysis from a broad perspective. As the usual unit costs were applied for the formal services, ‘shadow price’ was used for the informal care. CEA used the differences in costs and outcomes (palliative care outcome scale (POS‐8) and Zarit Burden Inventory (ZBI)) between baseline and follow‐up at 12 weeks. Total costs for 12 weeks measured at follow‐up were lower in the fast‐track intervention group than usual care group by GBP 1789 (95% CI: GBP ‐5224 to GBP 1902); converts to GBP 2424 (GBP ‐7077 to GBP 2577) in 2018. When inpatient care and informal care were excluded, mean service costs for 12 weeks were GBP 1195 lower for the intervention group (95% CI GBP ‐2916 to GBP 178); converts to GBP 1619 (GBP ‐3950 to GBP 241) in 2018. Cost‐effectiveness planes showed that 33.8% of replications for POS‐8 indicated that patients in the intervention group had lower cost and better outcomes than in the control group, and 54.9% had lower cost but worse outcomes. For ZBI, 47.3% of replications showed lower costs and better outcomes while 48% indicated higher costs and better outcomes.

McCaffrey 2013 estimated incremental net monetary benefit (INMB) and cost‐effectiveness acceptability curves (CEACs) for one extra day at home in an RCT among patients with mixed cancer and non‐cancer diagnoses with complex or unstable symptom management and high care needs in Australia. McCaffrey 2013 provided services across multiple settings. Data on resource use were prospectively collected and costed including: days at home, specialist palliative care service use, acute hospital and palliative care unit inpatient days, and outpatient visits. Intervention costs were calculated based on staff administration, travel and direct patient contact time, overheads and consumables. The analysis was conducted from a healthcare provider perspective and bootstrapping was used to calculate the confidence intervals around INMB and CEACs. Total costs were AUD 6452 (95% CI AUD 4469 to AUD 8586) (converts to GBP 5750 (95% CI GBP 3983 to GBP 7652) in 2018) in the HSPC group and AUD 5425 (95% CI AUD 2404 to AUD 8531) (converts to GBP 4835 (GBP 2142 to GBP 7602) in 2018) in the control group. The increment costs between the two groups was AUD 1027 (95% CI AUD ‐2612 to AUD 4738) (converts to GBP 915.22 (95% CI GBP ‐2327.71 to 4222.32). When the INMB of one more day at home was compared with varying threshold values, HSPC was preferred to usual care beyond AUD 1068. Sensitivity analyses with different inclusion ranges of costs (using hospital inpatient costs only and excluding high cost outliers) indicated that HSPC was preferred above AUD 2547 (converts to GBP 2270 in 2018) and AUD 846 (converts to GBP 754 in 2018). It was concluded that HSPC had a potential to be cost‐effective, especially in trials with longer follow‐up. The meaning of the threshold value for one extra day at home remains for future research.

Farquhar 2014 and Farquhar 2016 reported the cost‐effectiveness of the Breathlessness Intervention Service (BIS), a multidisciplinary complex intervention underpinned by a palliative care approach for patients with advanced cancer and advanced non‐malignant disease separately. The BIS was a model of HSPC where service provision traversed multiple settings in the UK. In Farquhar 2014, data from patients with advanced cancer were analysed from a societal perspective by including costs of informal care. Total health and social costs, including informal care for eight weeks prior to the baseline assessment, in the HSPC group were GBP 6137 (SD GBP 6099) which converts to GBP 6952 (GBP 6909) in 2018 and GBP 5461 (SD GBP 6099) which converts to GBP 6186 (GBP 6909) in 2018 for usual care. Costs between baseline and follow‐up at two weeks were GBP 794 (SD GBP 866) which converts to GBP 899 (SD GBP 981) in 2018 for HSPC and GBP 1121 (SD GBP 1635) which converts to GBP 1270 (SD GBP 1852) in 2018 for usual care. Intervention costs for HSPC were GBP 119 (SD GBP 62) which converts to GBP 135 (SD GBP 70) in 2018. Total costs were GBP 354 lower for HSPC (95% CI: GBP ‐1020 to GBP 246) which converts to GBP 401 (95% CI: GBP ‐1155 to GBP 279) in 2018 and incremental QALY‐gain was 0.0002 years (95% CI, −0.001 to 0.002), after controlling for baseline. The chance of HSPC being lower in total costs and providing better outcomes in terms of reduced distress due to breathlessness was 80.9% according to cost‐effectiveness planes and 16.4% for higher costs and better outcomes. It was 50.9% for the chance of HSPC being lower in total costs and greater in QALY, and 11% for higher costs and a greater QALY gain.

An NHS perspective was taken in the analysis of data from the UK study of patients with advanced non‐malignant disease in Farquhar 2016. Total health and social costs for eight weeks prior to the baseline assessment in the HSPC group was GBP 1952 (SD GBP 3290) which converts to GBP 2211 (SD GBP 3727) in 2018 and GBP 3630 (SD GBP 5588) which converts to GBP 4112 (SD GBP 6330) in 2018 for usual care. Costs between baseline and follow‐up at four weeks were GBP 1371 (SD GBP 2948) which converts to GBP 1553 (SD GBP 3339) in 2018 for HSPC and GBP 659 (SD GBP 1253) which converts to GBP 746 (SD GBP 1419) in 2018 for usual care. Intervention costs for HSPC were GBP 156 (SD GBP 80) which converts to GBP 177 (SD GBP 91) in 2018. With adjusting for baseline, total costs were GBP 799 higher for HSPC (95% CI: GBP ‐237 to GBP 1904) which converts to GBP 905 (95% CI: GBP ‐268 to GBP 2157) in 2018 and the HSPC group gained 0.003 extra QALYs (95% CI: –0.001 to 0.007). A cost per QALY for HSPC was GBP 266,333 (converts to GBP 301,692 in 2018). The chance of HSPC being lower in total costs and greater in QALYs was 7% according to cost‐effectiveness planes. There was an 86.5% likelihood of HSPC being higher in total costs and greater in QALY gain. The HSPC intervention appeared to be cost‐effective among patients with cancer but not among those with non‐malignant disease.

Mendoza‐Galindo 2018 (abstract only) compared resource use and costs between the early palliative care (EPC) group and usual care in patients with cancer diagnoses in Mexico. The study involved an outpatient model of HSPC. Resource use assessed included number/days of hospitalisation and emergency room (ER) visits as well as their cost. The number of ER visits in the EPC group was 39 while that in the control group was 50 (P = 0.074). There was also no difference in the number of hospitalisations (48% versus 51%) and days of hospitalisation (78 versus 90 days; P = 0.808) among both groups. Median cost associated with ER visits were non‐significantly lower in the EPC group (USD 21.99: converts to GBP 16.97 in 2018) compared to usual care (USD 46.35: converts to GBP 35.76 in 2018) (P = 0.081). The authors further reported lower median cost of hospitalisation days in favour of EPC (USD 167.57: converts to GBP 129.30 in 2018) compared to usual care (USD 295.05: converts to GBP 227.66 in 2018) (P = 0.015).

Ma 2019 assessed resource use and operating costs between an early palliative care intervention and usual care for patients in the ICU setting in USA. It was an inpatient consult model of HSPC. Resources used were extracted from patients' electronic medical records, including mechanical ventilation, vasopressors, haemodialysis, tracheostomy, cardiopulmonary resuscitation, ED visits, hospital readmission, hospital duration and ICU duration. Early palliative care patients had fewer ventilator days (median 4 versus 6; P = 0.042), tracheostomies performed (1% versus 7.8%; P = 0.035), postdischarge emergency department visits (1.3% versus 12.5%; P = 0.007), days on mechanical ventilation (median (IQR) 4 (3 ‐ 7) versus 6 (3 ‐ 13); P = 0.042) and hospital readmissions (17.3% versus 33.3%; P = 0.0024). There was no difference between the intervention and control group in ICU length of stay (median 5 versus 5.5 days), numbers on mechanical ventilation (53.6% versus 56.9%; P = 0.64), numbers on vasopressors (48.5% versus 50%; P = 0.83), days on vasopressors (median 3 versus 3; P = 0.91), numbers on haemodialysis (15.5% versus 23.5%; P = 0.15), numbers receiving cardiopulmonary resuscitation (5.2% versus 6.9%; P = 0.61) and hospital length of stay (median 10 versus 11 days). Analysis of operating costs was conducted though lacking in statistical power to detect the difference. Intervention patients had lower medical ICU (USD 9860 (converts to GBP 7608.08 in 2018) versus USD 15,660 (converts to GBP 12,083.42 in 2018); P = 0.004) and pharmacy costs (USD 3430 (converts to GBP 2646.62 in 2018) versus USD 5850 (converts to GBP 4513.92 in 2018); P = 0.016) per patient compared with the control group. However, the total operating cost per patient was not different between intervention and control group (USD 37,310 (converts to GBP 28,788.78 in 2018) versus USD 45,790 (converts to GBP 35,332.04 in 2018); P = 0.14). An estimated USD 880 (converts to GBP 679.02) of the intervention group’s per patient total operating cost was due to the added cost of the palliative care consultation.

Synthesis of nested or embedded qualitative studies that explored stakeholders' views and experiences of HSPC

Ten studies with a total of 322 participants (245 patients, 20 carers, 9 HSPC team members, 29 physicians (including oncologists), 14 oncology nurse practitioners, one consultant in interstitial lung disease, one clinical nurse specialist in interstitial lung disease, one community matron, one community palliative care nurse and one general practitioner) also had qualitative components that were used to explore stakeholders' views and experiences of HSPC (Bajwah 2015; Farquhar 2014; Farquhar 2016; Hopp 2016; Veron 2018 (linked to Janssens 2019); Lowther 2018 (linked to Lowther 2015); Maloney 2013 (linked to Bakitas 2009); Giovannetti 2018 (linked to Solari 2018); Talabani 2017 (linked to Brannstrom 2014); Wallen 2012) (see Table 16 under Additional tables). The number of patients interviewed by Wallen 2012 was unclear. However, a study (Slota 2014 linked to Wallen 2012) reporting the same data by the authors stated that 34 patients were involved in the qualitative analysis.

Data collection was mainly through semi‐structured interviews. However, Slota 2014 (linked to Wallen 2012) collected their data using open‐ended questions on a questionnaire, while Hopp 2016 qualitatively reviewed clinical records. Approaches to data analysis in these studies included content analysis, framework analysis and thematic analysis. Slota 2014 (linked to Wallen 2012) stated that they used thematic analysis while another study that reported the same data by the authors stated that they used transcript‐based analysis. Bajwah 2015 reported using a constant comparison approach within framework analysis, while Hopp 2016 did not state their approach.

Four studies were HSPC models involving service provision across multiple settings (Farquhar 2014; Farquhar 2016; Maloney 2013 (linked to Bakitas 2009); Wallen 2012), and another four were hospital outreach services (Bajwah 2015; Talabani 2017 (linked to Brannstrom 2014); Veron 2018 (linked to Janssens 2019); Solari 2018). Lowther 2018 (linked to Lowther 2015) was an outpatient HSPC model while Hopp 2016 was an inpatient consult model. Data from the studies were synthesised into two themes: valued components and challenges to HSPC provision.

Participants valued the patient and family‐centredness of the HSPC intervention as it helped to address the varied needs of patients and their unpaid caregivers/families. Benefits described included better symptom control, effective communication and shared decision‐making, psychosocial support and coping, respectful and compassionate care, supporting role maintenance and empowerment, reduced isolation, and improved use of devices. HSPC facilitated effective communication and shared decision‐making as patients and their unpaid caregivers/families had control over the care the patient received. They were able to ask questions, they were listened to and were able to receive the support they needed. Shared decision‐making and the psychosocial support provided as part of HSPC was therapeutic for patients and their unpaid caregivers/families, and also reassured them that they were not alone. Patients particularly valued services they received in the secure environment of their homes, and the involvement and support of their families. In addition to the care delivered, the process of delivery of care was also considered to be important. For instance, patients and their unpaid caregivers/families noted that the palliative care professionals were approachable, attentive and supportive. HSPC further facilitated care planning and the discussion of advanced care plans.

Although HSPC was viewed favourably by participants in these studies, there was also evidence that some participants questioned its usefulness. For instance, in Veron 2018 (linked to Janssens 2019), there were mixed reactions among advanced COPD patients about the value of the HSPC intervention. Authors described poor recollection of the HSPC consultation by patients who tended not to consider themselves to be sick. They ascribed their functional limitations to health problems other than COPD. Patients in this study avoided talking about the future and end‐of‐life issues and wanted to focus on the present. Also in Hopp 2016, participants expressed concerns that HSPC might prevent them from receiving more aggressive interventions and many did not want to discuss advanced directives.

Patients and their unpaid caregivers/families found the information provided during the HSPC intervention to be useful, as it ensured a better understanding of illness and treatment options. Patients and their unpaid caregivers/families valued the multidisciplinary nature of the HSPC team and their specialist expertise. Healthcare professionals such as oncologists tended to describe better patient care resulting from integration of palliative care with oncology at the time of diagnosis of advanced cancer.

Challenges to HSPC provision in these studies were identified, including lack of referral to HSPC by other health professionals, perception of palliative care as being synonymous with imminent death, lack of willingness to engage with palliative care, organisational barriers (e.g. insufficient services) and issues with the experimental study design (e.g. inadequate length of the HSPC intervention).