Managing uncertain recovery for patients nearing the end of life in hospital: a mixed-methods feasibility cluster randomised controlled trial of the AMBER care bundle

Clinical uncertainty has been defined as the inability to determine the meaning and significance of illness-related events [1]. It occurs when health professionals are unable to predict outcomes accurately due to insufficient information. Evidence suggests that in the last 30 days of life, the combination of deteriorating health and clinical uncertainty is highly distressing for patients in hospital and their families [2, 3]. This is amplified when discussions about their situation and preferences for care and location of death do not occur. Most people (67–80%) want to be informed about their poor prognosis [4]. Research, however, has identified that discussions about prognosis rarely occur [5]. This increases the likelihood of hospital deaths and also leads to poor care satisfaction, mistrust and loss of confidence in health professionals [6‐9] and may lead to complaints [10]. Clinical uncertainty also impacts the clinicians’ confidence and their practice. Health professionals frequently struggle with uncertainty, which can result in overtreatment or over-investigation [11], lack of communication with patients about their future [12, 13], and increased care costs [14].

In recent years, complex interventions [15, 16] aimed at improving the care of patients who may be approaching the end of life have become more common [17, 18]. In 2010, the AMBER care bundle was developed to improve care in the acute hospital setting for patients who are deteriorating, clinically unstable with limited reversibility, and at risk of dying in the next 1 to 2 months [19]. This was subsequently amended to be at risk of dying during their current episode of care despite treatment. The algorithmic intervention of the AMBER care bundle is designed to encourage health professionals to work with patients and families to develop and document a clear medical plan, including consideration of anticipated outcomes, cardiopulmonary resuscitation and escalation status, while acknowledging the uncertainty. This plan is revisited daily and encourages regular communication with the patient and their family regarding treatment plans, place of care and any other concerns. The bundle was designed to work alongside active medical care when uncertainty about the outcome remains.

A recent non-randomised comparative study of the AMBER care bundle with standard care, conducted in the UK, identified mixed findings. In comparison to similar patients in the control group, the use of the AMBER care bundle was associated with shorter lengths of hospital stay, more frequent discussions about prognosis between health professionals and patients, and higher awareness by patients of their prognosis. The clarity of the information provided, however, was rated lower by this group compared to those in the control group [20]. Qualitative research among health professionals has identified that the AMBER care bundle was often used as a tool to label or categorise patients, and indirectly served a symbolic purpose in affecting the behaviour of individuals and teams. Participants described the importance of the training associated with the intervention but reported that adequate exposure to the intervention, and the learning, varied [21].

Clinical equipoise, therefore, still exists in relation to the AMBER care bundle. A robust comparative evaluation of the intervention compared to standard care is, therefore, needed. The UK Medical Research Council’s guidance on the development and evaluation of complex interventions [15] and the Methods of Researching End-of-life Care (MORECare) statement [16] both recommend a feasibility study before full evaluation. Feasibility studies, now more common in palliative and end of life care [22], enable researchers to identify problems that might undermine the acceptability and delivery of the intervention or the conduct of a fully powered trial [23, 24]. Researchers are then potentially able to remedy problems with the intervention, trial design, or conduct by returning to the development phase, rather than proceeding to a full trial. This has important implications for the efficient use of resources, ensuring they are not directed to studies that produce a null result due to an unfeasible study design [25]. Moreover, it is also unethical to run a full trial before running a feasibility study.

In this paper, we report on the feasibility of conducting a pragmatic multi-centre cluster randomised controlled trial (RCT) of a hospital-based complex intervention (the AMBER care bundle) that aims to better serve patients whose situations are clinically uncertain and where there is a risk that they will die during their hospital stay, versus standard care. Four feasibility objectives were specified:

This study was registered with the International Standard Randomised Controlled Trials registry (ISRCTN: 36040085). Favourable ethical opinions were obtained from the national research ethics committee for Camden and King’s Cross (16/LO/2010) and the Health Research Authority. National Health Service (NHS) research governance approvals were obtained from each participating study hospital.

We conducted a parallel cluster RCT with a 1:1 allocation ratio, employing convergent mixed methods, with the quantitative and qualitative data given equal importance. Data were collected sequentially and analysed concurrently. A cluster RCT design was chosen because the implementation and delivery of the intervention were at an organisational level, in this case, hospital wards, and not the patients. Cluster RCTs are used to avoid contamination between treatment groups [26, 27]. This study comprised the trial, an examination of patients’ clinical records, and focus groups with health professionals.

The study took place across purposefully selected general medical wards in four clusters, in this case, district general hospitals (DGHs) in England. DGHs are major secondary-care facilities that typically provide an array of diagnostic and therapeutic services to the local population. There are over 142 DGHs in the UK [28]. The four DGHs selected serve diverse populations, including those with ethnic diversity and material deprivation. The hospitals have different strengths and weaknesses in terms of their Care Quality Commission ratings [29].

Participant recruitment and implementation of the AMBER care bundle were limited to one or two general medical wards at each hospital site. Selection of study wards at each site was informed by heat maps that provided contextual information at a ward level on the number of deaths during and up to 100 days after admission. Additional data comprised the number of patients who died with an individualised approach to the last days of life care and the number of hospital readmissions prior to a patient’s death. Wards with the highest number of deaths per year were considered to be suitable for the study.

Hospitals were randomly assigned to the intervention or control arms at the level of the cluster via an independent clinical trials unit by randomly sequencing the order of randomisation and then randomising the sites in this order into fixed blocks of two. Research nurses collecting data from patient participants were not masked to the group allocation. Quantitative analyses were performed, masked for the group allocation.

Research nurses identified patients (or their proxies) in the intervention and control wards daily who fulfilled the eligibility criteria aligned with those of the AMBER care bundle i.e. patients:

Research nurses scanned the hospital ward whiteboards to identify potential patient participants, who were then discussed with the clinicians to confirm their suitability for the study. All participants were considered to have mental capacity unless this was established otherwise, and all practicable steps were taken to enable individuals to decide for themselves whether they wished to participate. Potential participants’ level of capacity was discussed with referring clinicians to identify those with possible impaired capacity and to anticipate the likely consent procedure. Capacity was established in the initial meeting with the patient using the Mental Capacity Act (MCA) four-step process [30]. It was assessed whether the individual can:

Health professionals from study wards and the research nurses were invited via posters to participate in one of the four focus groups representing the study wards. Of those who expressed interest, we attempted to recruit a range of health professionals with different levels of experience. Written informed consent was obtained from health professionals prior to the focus groups taking place. Consent was obtained at the end of the focus groups from any health professionals who joined late but wished to participate.

The analysis followed the Consolidated Standards of Reporting Trials (CONSORT) guidelines (Fig. 1) and was conducted in collaboration with the clinical trials unit. Two statisticians (WG and RW), the chief investigator (JK), and the health economist (DY) were blind to the randomisation. Data were entered into predesigned Epidata databases [42]. In total 10% of the data were double entered and cross-checks were conducted. No discordance was detected for the candidate primary outcome measures (100% match for the IPOS subscale and howRwe) and very high accuracy for the rest of the questionnaires.

Since this was a feasibility study, a formal power calculation was not appropriate. Based on the information about the number of deaths and prior studies, we aimed to recruit 40 patients per study arm to provide us with sufficient data to test data collection forms and questionnaires, examine the appropriateness of candidate primary outcome measures, determine what would be the optimum data collection timing for a larger trial, and explore the acceptability of study procedures to patient participants. Any investigations of changes in study parameters were exploratory only.

Descriptive statistics on demographic and study variables were calculated as means, medians, ranges, standard deviations (SDs), and percentages (for categorical variables). No tests of significance were conducted. However, 95% confidence intervals, rounded up to one decimal place, were provided to indicate the precision of the estimates from the feasibility trial. The analysis of the IPOS data focused only on those participating patients with complete data for all IPOS patient and family anxiety and communication subscale items.

Economic evaluation is an emergent area in palliative care and ambiguity still surrounds best practice [39]. Procedures to inform the economic evaluation in the full cluster RCT protocol were reported, focusing on resource implications from health and social care, and societal perspectives. We aimed to make preliminary cost-effectiveness calculations (e.g. combining Client Service Receipt Inventory data on costs and the EQ-5D-5L score).

Responses to the five items in EQ-5D-5L were used to generate the index score for each patient. Theoretically, the index score ranges from 0 (death) to 1 (full health). Some EQ-5D-5L profiles were evaluated as below zero, implying that the individual considers their current quality of life as worse than death.

We had initially planned for recruitment to take place at each of the study sites for 3 months, with an expected average of seven participants being consented per month. However, recruitment was much slower than expected. Consequently, recruitment was extended to 10 months (June 2017 to March 2018). During this time, a total of 894 patients (130 in the intervention arm and 764 in the control arm) were screened for eligibility.

In the control arm, all patients on the study wards were screened by research nurses for potential eligibility against the inclusion criteria, which corresponded with the AMBER care bundle eligibility criteria. Patients, however, were deemed as being eligible only after confirmation by the clinical team. Subsequently, a member of the clinical team was required to explain to patients (or their relatives) their situation of clinical uncertainty and the purpose of the study. The most common reason preventing research nurses from approaching potential participants to ask for their informed consent was uncompleted clinical tasks, for example the failure to speak with family members about a patient’s current clinical situation. For the intervention arm, the identification of potential patient participants for inclusion in the study was guided by clinical decisions on whether patients were suitable to be supported by the AMBER care bundle. Since clinical teams assessed patients on the study wards to be supported by the AMBER care bundle on their daily handover meetings, the research nurses did not repeat this process. Although a larger number of patients were assessed for eligibility in the control arm of the study, the final number of patients who were eligible (n = 117) is very similar to those who were eligible in the intervention arm (n = 103).

Table 1 presents the number of patients screened and successfully recruited at each of the study sites. Only 1.9% (n = 8) and 8.9% (n = 28) of those screened in the control arm (both sites) were eventually recruited. The screening log for control site 1 provided detailed information on the characteristics of eligible and ineligible patients, and those who consented or refused. The log reported that of the patients screened who met the first eligibility criterion (deteriorating, clinically uncertain, and with limited reversibility), 55 (15.1%) did not meet the second criterion (at risk of dying during their current episode of care). Other sites provided only a limited range of information concerning potential recruits and control site 2 did not routinely keep a recruitment log. Recruitment rates were higher at the intervention sites, where 25.0% (n = 20) and 18.0% (n = 9) of those screened were recruited, respectively.

Clinicians and research nurses at the control sites reported challenges in identifying potential patients who fulfilled the eligibility criteria, particularly regarding the risk of dying during their current episode of care. Unlike health professionals at the intervention sites, they were not trained and guided in identifying these patients, as providing this education and support may have resulted in potential contamination of patient care. Since one of our study objectives was to examine critically how the study operated under field conditions, we reviewed patient recruitment in terms of the eligibility criteria over 4 months to assess the feasibility of recruitment. As a result, a pragmatic decision was made with the trial steering group to remove the ‘risk of dying’ criterion, focusing instead on just the first AMBER care bundle eligibility criterion, i.e. patients who are deteriorating and patients whose situations are clinically uncertain, with limited reversibility. A substantial protocol amendment was obtained from the NHS research ethics committee, in addition to local research governance permissions. We planned to monitor the effect of this change on recruitment, but control site 1 did not have the capacity to implement the revised recruitment strategy by the time approvals had been obtained. Control site 2 recruited eight more participants after this change.

Participants in both trial arms were predominantly white British and widowed, and most were either living comfortably or coping with their present level of income. The majority (64.6%, n = 42) of all 65 patient participants lacked mental capacity and therefore, proxy assent was required on their behalf. Control site 2, in an urban setting, had a more ethnically diverse sample profile compared to the other sites. There were differences between the trial arms (Table 2). In the control arm, most patients were men aged between 65 and 79 years with a cancer diagnosis, while in the intervention arm, the majority were women aged 80 years or older with a non-cancer diagnosis. The older age and non-cancer diagnoses of the patients in the intervention arm are likely due to the inclusion of two care-of-the-elderly wards at intervention site 1.

Reasons for the patient participants’ admission to hospital included shortness of breath, falls, and confusion. Out of 65 participants, 62 (95.4%) had an unplanned admission to hospital through an emergency department. They had a range of different illnesses (Table 3), and an average of 2.3 comorbidities (range 1–4) (Table 4), the most common being those associated with circulatory disorders.

Table 5 presents the levels of missing data and an exploratory analysis of patient primary outcomes for the IPOS subscale and the howRwe at each of the time points (mean and SD). The mean IPOS subscale score at the baseline was 13.3 (SD 4.8) in the control arm and 13.3 (SD 5.1) in the intervention arm, i.e. within the moderate range. This remained fairly consistent across the time points: 13.3 at 3–5 days (change of 0 from the baseline) and 10.3 at 10–15 days (change of − 3.0 from the baseline) in the control arm; 14.6 at 3–5 days (change of + 1.3 from the baseline) and 13.9 at 10–15 days (change of + 0.6 from the baseline) in the intervention arm. Although the howRwe scores changed across time points, it is not possible to comment on whether this is clinically significant because of the small sample size and the high rate of attrition between time points.

The descriptive statistics of service use showed that utilisation was within plausible ranges. Patients interviewed at 10 – 15 days, reported the use of investigations/tests and the informal care provided, but no health service use due to hospitalisation was reported between baseline and follow-up. Deriving EQ-5D index score was feasible for those who answered the questions on five dimensions (Table 6). Furthermore, we determined it was feasible to collect the data on health and social care service use, informal care provision and quality of life at baseline and at 10-15 days. Missing values in the data were not problematic (less than 9.0%). We decided not to calculate preliminary cost-effectiveness because attrition at 10-15 days reduced the number of paired samples available to twelve.

Patient participants considered their involvement in the study positively. Only one participant, who was in the control arm, did not want to complete the study questionnaire (no reason was stated). Some stated they were happy to participate due to the positive interaction with the research nurses, for example: ‘The research coordinator is very polite and explained everything about the study’ (Con2-014). Others were motivated by a sense of altruism, believing involvement would help others and improve services: ‘If you can help others, then it’s worth doing’ (Int2-007). A number of participants also encouraged other patients to take part in the study, reiterating that their involvement would ‘help others’.

In total, we conducted four focus groups with health professionals, one at each of the four study sites. Their views focused on the following issues: (i) the eligibility criteria for the AMBER care bundle and its implications for patient eligibility in the study, (ii) considerations of study settings and processes, and (iii) the impact of the feasibility study on research nurses. Details of the participants in the four focus groups at each of the study sites are presented in Table 7. Themes and illustrative quotes are presented in Table 8.

Referring clinicians and research nurses require that the eligibility criteria for a clinical trial to be clear and unambiguous. In the present study, the criteria created a number of sampling challenges outlined below.

First, the AMBER care bundle eligibility conditions operated as the eligibility criteria for the feasibility trial, which referring clinicians and research nurses found confusing. Beyond a patient being identified as deteriorating, clinically unstable with limited reversibility, patients were also required to be at risk of dying during their current episode of care, despite treatment. The combined evidence from the screening logs and the views of health professionals in the focus groups highlighted that the prognostic element of the criteria was a major impediment when identifying and recruiting potential patient participants. Whilst this finding was germane to both trial arms, it was more pronounced in the control arm due to lack of training and confidence in identifying potential patients. Prognostic models vary in levels of sophistication, ranging from clinical intuition to more intricate multivariate statistical models that combine multiple factors to yield an assessment [48]. If risk of dying is to be retained as an eligibility criterion, there is a continued risk of two sampling biases being present due to the unknown (or the inconsistent) manner in which health professionals currently interpret risk. Firstly, there is the unpredictable and often unreliable identification of potential participants within and across study sites, and secondly their exclusion. The findings from the health professional focus groups and analysis of the recruitment logs suggest both biases were indeed present. Although models to enhance the identification of dying patients using prognostic models are improving for patients with cancer [49‐51], there is far less consensus on methods to assess patients with non-malignant conditions, which are more common in studies of this nature [52, 53]. If similar complex interventions are to be evaluated, the subjectivity in prognostication must be avoided and greater emphasis placed on objective clinical indicators, for example, poor performance status scores, the presence and severity of cognitive impairment, weight loss, and dysphagia.

Second, equipped with the participant information sheets, clinicians were required to introduce the study to potential patient participants and its relevance to them in relation to their clinical situation. Some clinicians, especially at the control sites, reported they lacked the confidence and skills to talk openly with patients about their circumstances. This challenge was exacerbated by research nurses’ reports that while convinced that some patients identified were ideal for the study, their views were challenged by clinical colleagues who disagreed on their suitability, stating that they were unclear that they were at risk of dying during the admission. This inadvertent gatekeeping represented an important barrier to recruitment. In the clinical care of people approaching the end of life, patients value their autonomy in decision-making. This also applies to research participation, where the opportunity to help others, and to be heard, must be respected [54]. Preventing this, whether intentionally or unintentionally, may violate the ethical principle of fairness [55]. Future studies should test methods that train health professionals in conducting difficult conversations about introducing studies of this nature to potential patient participants, whilst being mindful that this training does not contaminate the study by corresponding too closely to the intervention to be tested.

Third, it was challenging to ensure homogeneity in how patients were identified at and across sites. Frequent staff turnover (notable at one site) may lead to inconsistencies (and potentially, bias) in the way potential patients are identified and recruited. It proved challenging for the research site’s principal investigators and researchers to identify whether the criteria were systematically applied and to track reasons for non-participation. Such tracking is invaluable, alerting researchers to the need to amend recruitment approaches before the study progresses too far and a vital pool of potentially eligible patients are inadvertently excluded. Related to this, accurate reporting of the number and characteristics of patient participants successfully recruited, and data on non-participants, greatly assists in the identification of possible sample bias that may compromise a study’s validity [56, 57]. We adopted the CONSORT and Transparent Reporting of Evaluations framework [58], aiming to report clearly and transparently the selection of the study sample in relation to the study’s eligibility criteria, the characteristics of participants and non-participants, and refusals. However, only one study site provided a detailed screening log that adequately met this requirement. Without this information, not only is it more challenging to identify and correct instances of misinterpretation of study eligibility criteria and to manage potential gatekeeping, but it also prevents the wider research community from understanding potential threats to the internal validity of studies examining similar issues. During the planning stage of a study, researchers need to be transparent about the resources required for the screening process and require study sites to record screening information. Since there is currently no agreed ethical standard for recording non-identifiable information within screening logs [59], training research staff, or delegated individuals, at each study site to collect minimum patient-based data may be helpful.

A number of important findings are evident from the process of seeking informed consent from patient participants. The participation information sheets and consent sheets were developed in concert with our patient and public involvement members and subsequently approved to the satisfaction of the research ethics committee and the UK Health Research Authority. However, we discovered they were not always well suited to their intended audience: those who were older, often frail, and very unwell. The research nurses highlighted that the documentation was lengthy, too detailed, and complex, a result of the need to include information and contingencies for those patient participants who might lose capacity during the study. Thus, some potential patient participants were discouraged from enrolling in the study. Whilst research has focused on the ethical requirements of dementia-related research, where important lessons can be learned [60], little guidance currently exists for developing study documentation for end-of-life care studies that are adequately detailed to satisfy potential participants and research ethics committees alike [61].

Future studies in this area should consider using a briefer version of the participant information sheet when first approaching potential patient participants. If they display interest in the study, then the full version of the participant information sheet could be provided to them that includes information on transparency in accord with the recently introduced General Data Protection Regulation and information on the legal basis for data-processing.

Related to this, the process of recording informed consent also requires consideration. According to the Council for International Organizations of Medical Sciences and the World Health Organization, it is permissible for researchers and research ethics committees to consider modifying consent procedures as long as they preserve as much of the informed consent process as possible to enable participants to understand the general nature of a study and to make a meaningful informed decision whether to participate [62]. The criteria for considering changes include: (i) the research would not be feasible or practicable to carry out without modification, (ii) the research has important social value, and (iii) the research poses no more than minimal risks to participants. We believe studies like ours should be permitted to test alternative methods of obtaining consent to evaluate their acceptability and utility from the perspectives of potential participants.

When designing this feasibility study, we deliberately made use of a cluster design to minimise the potential of study contamination [26, 27, 63] associated with the movement of health professionals acquainted with the AMBER care bundle to a control ward, which may influence the care given to patients. This represented an improvement in the design from our previous comparative evaluation of the AMBER care bundle [34]. We were also mindful not to select study sites where similar interventions were in place. This is because changes to the standard or usual care during a clinical trial could impair the validity of the study. Although control site 1 did experience a relatively frequent turnover of medical staff, to the best of our knowledge no health professionals were familiar with the AMBER care bundle from having worked elsewhere. However, at control site 2, we become aware of a change in clinical practice associated with subtleties in maturation, or naturally occurring changes [64] in health professionals’ clinical practice resulting from them becoming more familiar with the concept of clinical uncertainty, which was not accompanied by any formal training of an intervention for patient and family care. Due to the small number of participating patients in both arms of the study, it was not possible to quantify the effect of this change accurately.

We examined whether relevant clinical outcomes can be measured using instruments that could be easily completed by unwell patients whose recovery is clinically uncertain [65]. All participants (n = 65) who provided consent or proxy assent successfully completed the baseline measures. Overall, the levels of missing data for self-reported outcomes and those provided by proxies were very low for both candidate primary outcome measures. However, the howRwe is a patient-completed experience measure so that it cannot be used for those who lack adequate mental capacity. This severely restricted the number of participants who were able to complete this measure. Nevertheless, the findings indicate that data collection was generally possible. We also now believe that utilising health resources for the follow-up could be replaced by accessing patients’ medical records, assuming all the patients stay on wards. The costs associated with care service use would then be obtained using unit costs for each service item and opportunity costs (e.g. a minimum wage).

We have demonstrated that patient participants were generally very positive about being involved in this feasibility study. This continued for all those who remained in the trial until the second follow-up at 10–15 days. Within this study, we have refuted legitimate concerns about engaging with what could be perceived as vulnerable patient populations at the end of life [66‐68]. This study, therefore, demonstrates that when ethical and pragmatic decisions are made in relation to study design, combined with highly sensitised research nurses and researchers, the voice of patients can be heard.

There are a number of study limitations associated with this feasibility study beyond those already discussed. First, guided by the AMBER care bundle development team, we used heat maps to identify wards with the highest number of annual patient deaths. Consequently, we did not include wards with similar specialties across the trial arms. This resulted in a case mix of patients that was quite different between the arms. The effect of this was most pronounced with the inclusion of care-of-the-elderly wards, which skewed the age balance across the trial arms. The mean age of the participants in the intervention arm was higher than that in the control arm. Future studies should not base the selection of study wards solely on the number of deaths per ward, and should consider other important factors, for example, ward specialty, the potential for active engagement of ward staff, and the presence of principal investigators on the ward.

Second, we are mindful that this study represented a feasibility study with no requirement for a formal power calculation since effectiveness was not being evaluated. Based on available data of deaths on wards, or within 100 days of discharge, we estimated that over the period of the feasibility study, we would be able to recruit 40–45 patients in each arm. However, we recruited 65. This reduced number of patient participants has some implications for fulfilling the key objectives of the feasibility study, specifically, understanding how best to recruit, examining study participant retention, testing data collection tools for the trial and determining what would be their optimum timing in a larger trial, and examining the acceptability of the overall trial. Additionally, the relatively small number of clusters included in the study meant that we were not able to calculate the intra-cluster correlation coefficient required for a future trial. Feasibility studies adopting this trial design should consider extending the number of clusters.

Third, the main reason for the loss to follow-up was due to participants being discharged from the study wards, which was evident in both arms of the study. Since we collected data from participants only when they were on the study wards, we were not able to continue data collection after their discharge. Future similar studies should consider a design that either aims to recruit patients at an earlier point in their hospital admission or permits follow-up after discharge. A potentially more appropriate commencement point for recruitment within a hospital could be when patients are in an acute medical unit. Decisions regarding patients’ further treatment and care within a hospital often take place in the acute medical unit, and there is often clinical uncertainty at this time. This would allow data to be collected for a larger number of patient participants at the third time point (10–15 days), who would otherwise have been discharged.