Background Health coaching services could help to reduce emergency healthcare utilisation for patients targeted proactively by a clinical prediction model (CPM) predicting patient likelihood of future hospitalisations. Such interventions are designed to empower patients to confidently manage their own health and effectively utilise wider resources. Using CPMs to identify patients, rather than prespecified criteria, accommodates for the dynamic hospital user population and for sufficient time to provide preventative support. However, it is unclear how this care model would negatively impact survival.
Methods Emergency Department (ED) attenders and hospital inpatients between 2015 and 2019 were automatically screened for their risk of hospitalisation within 6 months of discharge using a locally trained CPM on routine data. Those considered at risk and screened as suitable for the intervention were contacted for consent and randomised to one-to-one telephone health coaching for 4–6 months, led by registered health professionals, or routine care with no contact after randomisation. The intervention involved motivational guidance, support for self-care, health education, and coordination of social and medical services. Co-primary outcomes were emergency hospitalisation and ED attendances, which will be reported separately. Mortality at 24 months was a safety endpoint.
Results Analysis among 1688 consented participants (35% invitation rate from the CPM, median age 75 years, 52% female, 1139 intervention, 549 control) suggested no significant difference in overall mortality between treatment groups (HR (95% CI): 0.82 (0.62, 1.08), pr(HR<1=0.92), but did suggest a significantly lower mortality in men aged >75 years (HR (95% CI): 0.57 (0.37, 0.84), number needed to treat=8). Excluding one site unable to adopt a CPM indicated stronger impact for this patient subgroup (HR (95% CI): 0.45 (0.26, 0.76)).
Conclusions Early mortality in men aged >75 years may be reduced by supporting individuals at risk of unplanned hospitalisation with a clear outreach, out-of-hospital nurse-led, telephone-based coaching care model.
- urgent care
- Machine Learning
- patient support
Data availability statement
Data are available upon reasonable request. All data requests should be directed to Health Navigator via InformationGovernance@hn-company.co.uk.
This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/.
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.
WHAT IS ALREADY KNOWN ON THIS TOPIC
Alternative models of care to reduce the strain on emergency services and improve health outcomes are needed globally. Community-based and patient-centred health coaching has been evidenced to improve patient experience, but not to reduce care demand or improve health outcomes.
It is hypothesised that identifying patients with amenable care in advance of a clinical crisis and supporting them to utilise healthcare resources proactively could improve the overall impact of patient-centred interventions like health coaching.
WHAT THIS STUDY ADDS
Data from a multicentre randomised controlled trial, involving 1688 participants targeted via a clinical prediction model, suggested that health coaching has no significant effect on overall mortality.
Further exploratory analysis found a significantly lower mortality rate in men aged >75 years (a preplanned subgroup definition at 50% quantiles).
The sensitivity analysis, excluding a site unable to adopt a clinical prediction model, suggested a larger mortality reduction for men aged >75 years.
HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
Predicting unplanned hospitalisation using routinely collected secondary care data, and supporting at-risk patients earlier with remote, anticipatory care could help save lives and address gender-related health inequalities.
The crowding of emergency departments (ED) is a major global issue having a detrimental effect on health outcomes by preventing access to timely care.1 2 The cause of ED crowding is complex and multifactorial,3 but many emergency visits are avoidable4 5 and a small number of patients account for a disproportionately large amount of unplanned care.6 7
Health coaching is one of many proposed patient-centred interventions designed to empower patients to manage their own health and help mobilise their access to wider healthcare resources.8 9 Associations have been evidenced between care continuity and reduced demand10–12; thus, patient-centred approaches like health coaching are expected to, in turn, reduce care consumption. However, despite an abundance of health coaching literature demonstrating a positive impact on patient experience,13–15 evidence of it reducing care consumption and improving clinical outcomes is limited,16 and the quality of evidence varies.17
Moreover, it has been hypothesised that predictive approaches are required to effectively target health coaching interventions at the appropriate population in advance of a clinical crisis to achieve optimal impact from a patient, clinical and resource perspective.18 19 This is because frequent hospital users are a dynamic population, with many frequent users typically returning to expected demand within 12 months and others becoming high users.20 Existing research investigating health coaching on clinical outcomes such as mortality is restricted to targeting patients using prespecified criteria, without the use of predictive analytics.21–23
This study aimed to investigate whether identifying patients in advance of them seeking emergency care using a prediction model and supporting them with health coaching could reduce future emergency admissions and ED visits. In this preliminary analysis, the impact of this intervention on mortality (which was studied as a safety outcome) was evaluated.
All methods and results have been reported in line with the Consolidated Standards of Reporting Trials (CONSORT) reporting checklist (see online supplemental appendix A).
Data were used from a multisite parallel two-arm randomised controlled trial (RCT) where CS-J (Senior Research Analyst, Nuffield Trust) was the chief investigator leading the evaluation protocol, and the intervention was delivered by Health Navigator (Integrated Research Application System project ID: 173319; and International Standard Randomised Controlled Trial Number: 12724749). The RCT was designed to evaluate the impact of a telephone-based health coaching intervention targeted to patients estimated to be at elevated risk of future unplanned hospital care via a clinical prediction model (CPM), where mortality was monitored as a safety endpoint (all RCT outcomes are clarified below). Recruitment started in August 2015 and concluded in 2019. The main trial analysis endpoints are not reported in this manuscript as complete participant-level linked datasets are not yet available. Interim datasets of hospital activity were used to provide insight into the health status of the trial population.
Study population and randomisation
Patient-level data were obtained from each site (York Teaching Hospitals, East Kent Hospitals University Foundation Trust, University Hospitals Derby and Burton, University Hospitals North Midlands, Royal Wolverhampton Trust, Mid and South Essex NHS Foundation Trust) and linked with NHS Spine to obtain both inpatient and out-of-hospital deaths. NHS Spine data, managed by NHS Digital, refer to a linked collection of local and national databases containing both a patient’s clinical and demographic information.24 All hospital patients with an emergency attendance or inpatient admission in the preceding 6 months were automatically screened for trial inclusion by a CPM, locally trained on 3 years of routine hospital data and designed to identify adult patients at risk of 3 or more unplanned bed days within 6 months. These GLMnet models were internally validated using cross-validation and achieved a discrimination area under the curve of 0.8 (more details and leading predictors can be found in online supplemental appendix B.1). The nurse-led manual screening assessed suitability for intervention. Participants unable to hear and communicate in English without a translator, or with a history of cognitive impairment or mental disorders were deemed unsuitable for the intervention (see online supplemental appendix B.2.1). The RCT invitation rate from the CPM was defined as the proportion of participants categorised as suitable by the coaches at the point of medical record screening out of all those identified by the local CPM and screened by a coach. Consented patients were openly randomised using an automated random-sequence generator to either the intervention or control group using a 2:1 ratio, respectively (see online supplemental appendix B.2.2). Control patients were not contacted after participant consent was obtained.
Primary, secondary and safety outcomes for trial participants were collected for 2 years post-randomisation. Mortality was the safety outcome. The primary outcome was non-elective admissions and the secondary outcomes focused on a wider set of hospital activity metrics and patient outcomes (see online supplemental appendix C).
The trial was powered to detect a clinically meaningful reduction (12–15%) in emergency admission rates, with 90% statistical power (see online supplemental appendix B.2.3). Although the initial power calculations suggested an overall sample of 3000, local principal investigators and the Nuffield Trust mutually agreed to close recruitment due to no longer supporting the withholding of beneficial resources from control participants. This decision was informed by interim analysis investigating hospital activity and patient-reported impact.
Intervention participants received an initial face-to-face assessment meeting with one of Health Navigator’s health coaches, a registered health professional, to better understand their health and social care needs and establish any gaps in their routine care. The initial assessment promoted the development of a personalised care plan and was followed by a series of one-to-one telephone-based coaching sessions with the same coach. During these calls, patients received motivational guidance, support for self-care, health education, and coordination of social and medical services. Motivational guidance was informed by existing theories (eg, 25); components included demonstrating empathy, dealing with resistance, supporting self-efficacy and developing autonomy. No medical advice or treatment was delivered. Control participants received NHS standard care.
Data collection and management
The trial participants were linked to the NHS Spine data using their NHS number and date of birth. Patients were excluded from the analysis if they were unable to be linked, or if they had missing age, sex or deprivation.
This section reports the key statistical methods for interpretation; further methods are provided in online supplemental appendix D. The research protocol designed before trial delivery can be found in the corresponding Integrated Research Application System (IRAS) form and online supplemental appendix C. Please note that although mortality was specified as a safety endpoint before the trial with an inspection of subgroups defined by 25% and 50% quantiles, the modelling choice described in this section was proposed after the trial but before any analysis was performed on the NHS Spine data in R.
To compare patient demographic case-mix across treatment groups, summary statistics were provided for age, sex and Index of Multiple Deprivation (IMD; as a measure of social economic status) decile. Crude survival rates were compared across groups. The 2-year absolute risk reduction and associated number needed to treat (NNT) to prevent one death were calculated to provide a clinically interpretable measure of impact. Kaplan-Meier curves were also compared across treatment arms, and for age, gender and IMD subgroups. Age was divided into two categories using the population median for pairwise interactions, aged under 75 and 75 years and over. The IMD deciles were also divided into two categories: the 50% most deprived areas (IMD decile 5 and under) and 50% least deprived areas (decile >5).
Bayesian survival analysis was employed because prior distributions can increase statistical power and posterior probabilities have a valid interpretation without formal hypothesis testing. To help with interpretation here, posterior probabilities indicate the probability of a true association between the intervention and mortality in a population conditional on the observed sample and previous evidence; informative prior distributions reflect what previous research suggests, in this instance, the true population-level HR could be, and posterior distributions suggest the distribution of possible true HR values given previous research and the new data analysed.
A Bayesian-Weibull survival model26 was fitted with fixed effects for the intervention group (binary), the top 50% most deprived (binary), age (in decades, or above/below 75 years) and sex (binary), and a random effect (frailty) for the site. Although age was categorised for pairwise interactions, it remained continuous as a control variable. The random effect accounts for variation in the baseline hazard between sites and assumes proportionality. Previous studies evaluating similar interventions suggested little impact on mortality with high levels of uncertainty. Thus, the prior distribution biased the model towards a null effect (HR=1) of the intervention on mortality and assumed the HR would most likely lie between 0.2 and 5 (see technical details in online supplemental appendix D.1.1.1). First-order interactions between the intervention and each control variable were added into the model, and then second-order interactions between age and gender were investigated given the first-order interactions. Posterior probabilities are reported for all Bayesian models.
To investigate the potential impact of modelling choice, the analysis was performed again with non-informative priors and using a Cox proportional hazards model. As one of the sites did not have the technical readiness required for CPM deployment, sensitivity analysis was also performed without this site. The digitalisation of paper records at Mid-Essex Hospital Services NHS Trust only started in 2017, leading to limited historical patient information electronically feeding into the CPM.
To explain the findings, post hoc analyses were performed on the intervention participants using the health coaches’ data, primary care activity data for a subset of York participants and using the unlinked local hospital interim data. To gain insight into the social background of the intervention participants, proportions of men >75 years living alone were compared with rest of the intervention arm. To gain insight into their medical profiles, patient-reported clinical diagnoses were compared for men aged >75 years and the rest of the intervention arm. Comparisons could not be made to the control group due to the data generation process. To compare medical profiles between groups, International Classification of Diseases (ICD)-10 codes associated with inpatient stays were compared for men aged >75 years, comparing the distribution across Charlson Comorbidity Index (CCI) groupings27 and the most common ICD-10 codes associated with their most recent activity. To explore whether the intervention led to an increase in primary care utilisation, incidence rates of activity were compared across subgroups of interest. To explore whether the intervention led to an increase in secondary care utilisation, activity counts 6 months pre-randomisation and 12 months post-randomisation were provided.
The CONSORT flow diagram illustrates this study’s inclusion process in figure 1. The invitation rate from the CPMs, averaged across sites, was 35%.
The participant characteristics summarised in table 1 suggest a similar distribution of age, gender and deprivation status across both treatment arms. Both groups have a median age of 75 years (IQR=15 and 14), an even balance of genders (52.8% vs 52.5% female) and just under one-third of the group living in the top 50% most deprived areas (30.8% vs 29.3%). Aggregated data from the local sites suggest that both treatment groups had similar hospital utilisation 6 months pre-randomisation and were in hospital for similar conditions during this time (see table 1). The top 10 ICD-10 codes presented for each participant’s latest hospital admission pre-randomisation are provided in online supplemental appendix D, with an agreement of 5 out of 10. Trial participants across treatment arms also displayed similar time distributions since their latest hospital activity pre-randomisation (see online supplemental appendix D.5).
Kaplan-Meier curves and crude mortality rates across treatment arms and other participant subgroups suggested an overall reduction in mortality for intervention participants (see table 2 and figure 2, log-rank p=0.13), for male patients based in the 50% least deprived areas (Figure D.2 online supplemental appendix D, log-rank p=0.050), and for patients aged 75 years and over in the 50% least deprived areas (Figure D.3 online supplemental appendix D, log-rank p=0.037). Comparing crude mortality rates using Kaplan-Meier curves suggested that the largest reduction in mortality across treatment arms was for men aged >75 years (log-rank p=0.0011, NNT (95% CI)=8 (5, 27)). The difference in rates was statistically significant according to the log-rank test (see figure 3).
Statistical summary and modelling
Estimated HRs and credible intervals (CIs) for the Bayesian-Weibull survival models (with an informative prior) are reported in table 2. The posterior distributions of the estimated HRs associated with the intervention are illustrated in figure 4.
According to the adjusted results in table 2, the intervention group showed lower overall mortality compared with the control group (HR=0.82, 95% CI=0.62, 1.08), although this was not statistically significant. Male patients showed a strong reduction in mortality (HR=0.71, 95% CI=0.50, 1.01), whereas females showed little difference in mortality compared with the control (HR=1.02, 95% CI=0.67, 1.62). Among the two age categories, a significant reduction in mortality was only observed in patients aged 75 years and over (HR=0.72, 95% CI=0.52, 0.99). The strongest reduction in mortality was observed in male patients, aged 75 years and over (HR=0.57, 95% CI=0.37, 0.84), suggesting most of the impact is being driven by this participant subgroup. A stronger association with reduced mortality was identified for participants in the 50% least deprived areas (HR: 0.79, CI=0.56, 1.10) compared with those in the 50% most deprived (HR: 0.90, CI=0.55, 1.48), but a wide credible interval for those in the highly deprived areas suggests prominent uncertainty due to fewer participants.
The sensitivity analyses suggested consistent findings across the choice of prior and statistical model (crude reduction of 50.5% for men aged 75 years and over, non-informative prior HR (CI): 0.54 (0.35, 0.83) and Cox proportional hazards HR (confidence interval): 0.55 (0.37, 0.83), see table D.3 in online supplemental appendix D). However, stronger reductions in overall mortality rates for men aged 75 years and over were observed once the Mid-Essex site participants were removed (HR: 0.45; CI=0.26, 0.76 without Mid-Essex compared with HR: 0.57; CI=0.37, 0.84; see table D.4 in online supplemental appendix D). The post hoc analysis on intervention participants suggested that the mortality impact observed on men aged 75 years and over may not be explained by their living situation but comparisons with matching controls are required (see online supplemental appendix D.4.2.1). Both the patient-reported medical diagnoses in the intervention group and comparison of ICD-10 codes associated with inpatient stays suggested that men aged 75 years and over in the control group were not more unwell than those in the intervention group (see online supplemental appendix D.4.2.1 and D.4.1.4). A further exploration controlling for ICD-10 groupings in the Bayesian survival analysis on a participant level is required to explore whether they contribute to the mortality difference. Furthermore, an increased primary or secondary care utilisation was not observed for this participant subgroup (see online supplemental appendix D.4.2.2 and D.4.2.3).
For patients identified by a CPM to be at risk of future unplanned hospital activity, the findings suggested that health coaching was associated with a mortality reduction for men aged 75 years and over. Stronger associations with reduced mortality were estimated when a single site, unable to employ a CPM, was excluded.
The overall findings of this study align with previous evidence for health coaching for patients with multiple chronic conditions,22 cardiovascular disease21 and for elderly patients in managed care.19 Null effects on mortality were found in other studies however.20 23 24 Granular investigations of the association between the intervention, survival and participant subgroups are limited with most studies focused on the overall impact.19 20 22 A single study reported a greater association between the intervention and reduced mortality for those with more coaching sessions, and for those who were male, but the intervention was restricted to patients diagnosed with cardiovascular disease,21 while our study found a significant effect only in men over 75 years. It is difficult to understand the impact of using a CPM in previous literature, as they have not been combined with health coaching in the same way. A rule-based approach (aged over 65 years, lived in community and had at least one risk factor relating to hospital activity, multimorbidity, frailty or living situation) employed with a collaborative primary care-based intervention (the addition of a nurse and care assistant) in Illinois was shown to reduce 24-month all-cause mortality when baseline differences in morbidity, hospital use and education were controlled for, but similar crude mortality rates were observed between groups after randomisation (11% vs 9%) and the sample size was small.28
It is important to acknowledge the study limitations when interpreting the findings. First, the study was not initially powered for mortality as it was not a primary outcome for the trial. Nevertheless, a strong association was identified in a subset of the study population and Bayesian inference was employed to aid with clinical interpretability of the findings given the available data. Second, limited access to the full dataset and primary care data prevented the ability to control for care utilisation in the Bayesian models and ensure balanced medical case-mix across treatment groups. However, potential bias and imbalance across treatment arms were explored using sensitivity analyses. Any differences between the groups were also determined by chance and highly unlikely given the extended screening process and randomisation. Third, some patients were excluded from the study due to missing demographics or linkage problems with the Spine dataset, but these numbers were small. Fourth, the CPM invitation rate was lower than expected (averaged 1 in 3). This has been attributed to data missingness in model deployment and screening, and patients still showing as admitted after hospital discharge. After adjusting for these learning points from the RCT, the invitation rate has since improved to 1 in 2 (50%).
Impact and implications
This study suggests that a health coaching service targeted to patients using a CPM is associated with positive survival impact for male patients aged 75 years and over. The sensitivity analysis and existing evidence in this area indicate that the mortality impact for elderly men could be due to the intervention (1) identifying potentially vulnerable patients in a timely manner using a CPM,29 (2) being more proactive in enabling access to health services for those more vulnerable due to health inequalities30 and (3) ‘recruiting’ patients for extra support rather than ‘referring’ them onto other services.31 The study findings highlight health inequalities, driven by gender, which receive minimal attention in the literature.30 Further investigations using the full trial dataset, once accessible, are required to establish the core drivers of the mortality impact. Furthermore, these findings do not suggest that the intervention should only be implemented for this participant subgroup, as care consumption and patient outcomes are the prioritised endpoints.
This study has demonstrated that predicting individuals at risk of unplanned care within 6 months and supporting them with an out-of-hospital care model does not have a negative impact on survival. On the contrary, the exploratory analysis suggested it could have a positive impact on survival rates for men aged >75 years.
Data availability statement
Data are available upon reasonable request. All data requests should be directed to Health Navigator via InformationGovernance@hn-company.co.uk.
Patient consent for publication
This study involves human participants and received ethical approval from the NHS Research Ethics Committee and is registered with the National Institute for Health Research (Central Portfolio Management System reference: 19857) and Health Research Authority (Integrated Research Application System ID: 173319). Participants gave informed consent to participate in the study before taking part.
The authors would like to note the outstanding efforts of the health coaching nurses and other staff involved in the provision and development of this intervention who made this study possible, and of course the patients. Special thanks also go to the principal investigators at the each of the study sites for their substantial contribution to the delivery and success of the trial (in particular, for York: Dr Nigel Wells, George Scott and Fiona Bell Morritt; for East Kent: Dr Marc Farr; for Staffordshire: Dr Paddy Hannigan). The authors would also like to thank Jamie Davis for the natural language processing analysis performed to exchange free-text data to SNOMED codes as reported in the supplemental material. The authors would also like to thank Graham Prestwich, patient involvement and engagement lead at Yorkshire and Humber Academic Health Science Network for leading and delivering the patient involvement to supplement the presented analysis.
Handling editor Richard Body
Contributors LMB, BA, AN, TL-B and JW conceived the study. LMB and BA developed the study design and analytical approach, in consultation with other project team members. TL-B performed initial analyses and extracted the relevant data sources. BA performed the statistical analyses, which were supervised by LMB, AN and JW. LMB, SR and BA performed the supplementary descriptive analysis on local trial sites data. Data and coding scripts were quality assured by CS-J. LMB drafted the initial version of the manuscript. All authors contributed to the interpretation of the findings and reviewed and edited the manuscript for intellectual content. All authors approved the final version of the manuscript and agreed to be accountable for all aspects of the work.
JW, as the guarantor, was responsible for the overall content of this manuscript.
Funding Research and development contracts were agreed between Health Navigator and the local trial sites and their clinical commissioning group (CCG); thus, the trial was funded through NHS CCGs partaking in the randomised controlled trial. Named funders in this respect are Cannock Chase CCG, East Kent CCG, Mid-Essex CCG, South East Staffordshire and Seisdon Peninsula CCG, Stafford and Surrounds CCG, Vale of York CCG and Wolverhampton CCG. Any costs relating to the submission, approval and amendments made to the HRA, as well as management of the trial, were funded by Health Navigator.
Competing interests LMB, BA and SR are employed by Health Navigator as data scientists. AN is the head of analytics at Health Navigator. JW is the founder and chair of Health Navigator. Contractual agreements were in place between Health Navigator and East Kent Hospitals University NHS Foundation Trust for the initial data extraction and analysis by TL-B. Contractual agreements were in place between Health Navigator and Nuffield Trust for the design and oversight of the RCT analysis by CS-J. MS declared no conflicts of interest.
Provenance and peer review Not commissioned; externally peer reviewed.
Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.