Understanding the factors influencing participant engagement and adherence in exercise referral in the City of Manchester

Abstract

Objectives

To identify demographic, clinical, socioeconomic and referral-pathway determinants of engagement, early dropout and non-participation in a large metropolitan exercise referral scheme (ERS), and to assess whether machine learning (ML) methods provide additional explanatory value beyond conventional regression.

Methods

This retrospective cohort study analysed data from 11 909 adults (≥18 years) referred for ERS in Manchester, UK, between 27 January 2022 and 28 February 2025. Outcomes were programme adherence (completion of ≥12 exercise sessions), early dropout and non-participation. Multinomial logistic regression was used to examine predictors of outcomes. ML methods, including Random Forest models and k-means clustering, were applied to assess predictive performance, rank feature importance and identify participant subgroups.

Results

Of 11 909 referrals, 34.6% completed the programme, 34.2% declined to participate and 8.0% left early. Younger age, higher neighbourhood deprivation and psychosocial referral reasons (eg, loneliness, long covid) were associated with lower completion. Participants referred through outreach/social prescribing pathways were substantially more likely to remain in an ‘intends to participate’ state rather than complete, compared with those referred through clinical pathways. Random Forest models demonstrated good predictive accuracy, with early engagement indicators and deprivation emerging as the strongest predictors. Clustering identified a high-risk subgroup characterised by younger age, higher deprivation and low early attendance.

Conclusions

Engagement and adherence in ERS were strongly shaped by early engagement, deprivation, age and referral pathways. ML methods identified high-risk subgroups and reinforced the importance of early attendance. Targeted early support and ML-informed risk indicators may improve retention, particularly among younger and more deprived participants.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Exercise referral schemes (ERS) are widely used across the UK as a public health approach to increasing physical activity among people with, or at risk of, chronic conditions. However, despite reach, engagement and adherence remain inconsistent, and many schemes report substantial levels of drop-out. Existing research has explored socio-demographic and health-related factors that may influence participation, but findings have been mixed and often depend heavily on local context.

WHAT THIS STUDY ADDS

• This study provides the first large-scale metropolitan analysis of ERS engagement using machine learning (ML) alongside traditional regression. It identifies deprivation, sex, health conditions and early engagement as key predictors of adherence and highlights the pivotal role of referral pathways. Outreach routes appear to broaden access but also reveal an intention–behaviour gap. At the same time, some deprived wards show strong completion once participants engage, challenging assumptions about uniformly poorer outcomes in disadvantaged areas. The combination of ML and regression offers new insight into dropout risk and patterns of participation.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• The findings support the need for targeted interventions for younger males and individuals referred for psychosocial reasons, who appear most vulnerable to dropout. Strengthening triage processes, improving the quality of referrals and embedding warm handovers where referrers personally introduce participants to ERS staff to maintain engagement or structured follow-up mechanisms may help to convert initial intent into active participation. The analysis also suggests value in co-designing community outreach approaches with local partners to improve retention in deprived communities. Finally, the use of ML-based risk flags offers a practical tool for identifying participants who may require additional support, contributing to more personalised and equitable ERS delivery.

Introduction

Physical inactivity is a major global health risk, associated with a 20–30% increase in disease-related mortality.¹ Public health strategies increasingly prioritise interventions linking healthcare with structured physical activity. Exercise referral schemes (ERS) were introduced in the UK in the early 1990s to provide supervised activity programmes for individuals with medical conditions or risk factors.^{2 3} Despite widespread use, engagement and long-term adherence remain inconsistent, with dropout rates commonly exceeding 40%.⁴

Demographic, health-related and socioeconomic factors influence ERS participation, yet findings are mixed and often context-specific.^{5 6} Many evaluations use relatively small samples, often involving only a few hundred participants, limiting generalisability and highlighting the need for larger, whole-system analyses.^{2 4} National evaluations indicate that younger adults and residents of deprived communities are less likely to complete programmes.^{3 7} However, local schemes vary widely in delivery models, referral pathways and settings, making localised analyses essential. Traditional regression approaches capture some determinants of engagement but may struggle with the complexity of demographic, clinical and behavioural interactions.

Manchester represents an important case for investigation. The city has a diverse population of over 550 000, a young median age of 34 compared with the national average of 40.7 years⁸ and high levels of deprivation, with around a quarter of residents living in the most deprived neighbourhoods. These factors contribute to significant health inequalities, including elevated rates of chronic illness and mental health concerns.⁹ In response, the Physical Activity Referral Service (PARS) was launched in 2018 by MCRactive, in partnership with healthcare providers, community organisations and the local authority. Referral routes include general practitioners (GPs), physiotherapists, social prescribers and other health professionals, with programmes delivered across leisure centres and community venues. Despite investment, challenges around uptake and sustained engagement persist, and the drivers of these outcomes within this metropolitan context remain unclear.

Machine learning (ML) offers an opportunity to complement traditional statistical approaches by modelling non-linear relationships, identifying complex interactions, ranking predictor importance and detecting subgroups at heightened risk of dropout.¹⁰ Although applications in physical activity and health services research are limited, early evidence suggests its potential to inform targeted interventions and improve programme adherence.¹¹

Therefore, this study aimed to examine factors influencing engagement and adherence within Manchester’s ERS by integrating traditional statistical analyses with ML to identify predictors of adherence, explore early dropout and assess the influence of referral pathways on participant outcomes.

Methods

Study design and participants

This retrospective cohort study analysed Manchester Active ERS data collected between 27 January 2022 and 28 February 2025. Data were extracted from the ReferAll system, which records referrals, engagement and outcomes. Adults ≥18 years were included. Duplicate records (n=3025) and referrals to third-party providers without chronic conditions (n=5536) were removed, yielding 11 909 participants.

Outcome measures and predictor variables

Primary outcomes were adherence, dropout and non-participation. Adherence was defined as completing ≥12 exercise sessions (up to 24), consistent with Manchester Active’s operational criteria and broadly aligned with the traditional 12-week ERS model.^{2 12} Dropout was defined as leaving before completing the prescribed sessions, and non-participation was defined as declining engagement following referral. Referral status was originally coded into five categories; analyses focused on the three most frequent: Completed, Left Early and Not Participating.

Predictor variables included age (Office for National Statistics age bands), sex, neighbourhood deprivation (Index of Multiple Deprivation, 2019, quintiles), referral source and referral reason (eg, musculoskeletal, diabetes, mental health). Ethnicity was excluded due to >40% missing data.

Statistical analysis

Descriptive statistics summarised participant characteristics and outcomes. Group differences were assessed using χ² tests for categorical variables and t-tests/analysis of variance for continuous variables. Multinomial logistic regression examined predictors of outcomes, with ORs and 95% CIs reported (p<0.05).

Machine learning analysis

Supervised and unsupervised ML methods were used to complement regression analyses. Random Forest and Decision Tree models were used to predict outcomes, with performance assessed using accuracy, precision, recall and F1 scores. Random Forests were prioritised due to stronger performance and the ability to rank feature importance. K-means clustering identified participant subgroups based on demographic and engagement variables.

All analyses were conducted using routinely collected, anonymised service-level data provided by Manchester Active under a data-sharing and confidentiality agreement. Data handling complied with General Data Protection Regulation (GDPR, 2018) and institutional governance requirements.

Patient and public involvement statement

Patients or members of the public were not involved in the design, conduct, reporting or dissemination of this study.

Results

Programme uptake, dropout (Left Early) and completion

Between 27th January 2022 and 28th February 2025, 11 909 participants were referred to Manchester’s ERS (after exclusions; see online supplemental figure S1). Uptake, defined as accepting participation following referral, was high, with 57.8% of referred individuals engaging with the programme. Of these, 34.6% completed the programme, while 8% left early (classified as dropouts in this study). A further 34.2% declined to participate and were therefore categorised as non-participants, with a smaller proportion still engaged or awaiting consultation. Referrals were most frequently received from hospitals (38.7%), GP surgeries (25.2%), medical centres (20.7%) and outreach services (15.2%). Most participants (62.9%) lived in the most deprived quintile of neighbourhoods in England (Index of Multiple Deprivation (IMD), 2019), an area-level measure of relative deprivation, with fewer than 1% residing in the least deprived quintile.

Participants presented with a diverse clinical profile typical of an urban ERS population. Musculoskeletal (MSK) conditions, referring to disorders affecting bones, joints and soft tissues such as arthritis, back and/or knee pain and other movement-related problems, were the most common referral reason (39.2%), followed by metabolic conditions, including type 2 diabetes (6.6%) and weight management referrals (7.6%), and respiratory conditions such as chronic obstructive pulmonary disease (COPD) (7.3%) and asthma (2.8%) were also well represented. Cardiovascular conditions accounted for a smaller proportion of the cohort, including stroke/transient ischaemic attack (TIA) (3.1%) and previous myocardial infarction (1.3%). Psychosocial and mental-health-related referrals formed another notable subgroup, with primary mental health concerns (3.7%), loneliness (2.1%) and long covid (0.2%) contributing to the overall case mix. Collectively, this clinical profile reflects the multimorbidity and psychosocial complexity characteristic of densely populated and highly deprived urban settings.

Predictors of adherence and dropout

Multinomial logistic regression identified several demographic and clinical predictors of programme outcomes. Participants with MSK conditions, alongside those with type 2 diabetes, cardiovascular disease, asthma, stroke/TIA, COPD and a history of falls, were more likely to intend to participate but not complete, compared with completers (table 1). Younger participants and those referred for psychosocial reasons, such as long covid or loneliness, were less likely to complete (tables2 3).

Table 1. Odds of intending to participate by health condition.

Health condition	OR	95% CI	P value
Musculoskeletal conditions	4.07	(3.14 to 5.29)	<0.001
Type 2 diabetes	2.38	(1.59 to 3.57)	<0.001
Cardiovascular disease	3.39	(1.27 to 9.05)	0.015
Asthma	2.66	(1.48 to 4.79)	0.001
Stroke or TIA	2.27	(1.19 to 4.31)	0.013
COPD	1.71	(1.08 to 2.73)	0.023
Falls	1.68	(1.14 to 2.49)	0.009
Long covid	0.12	(0.04 to 0.34)	<0.001
Arthritis	0.41	(0.31 to 0.52)	<0.001
No weight management referral	0.27	(0.21 to 0.35)	<0.001
Reported loneliness	0.31	(0.19 to 0.52)	<0.001

Multinomial logistic regression results showing the likelihood of intending to participate relative to completing the programme, with completion as the reference category.

COPD, chronic obstructive pulmonary disease; TIA, transient ischaemic attack.

Table 2. Odds of leaving the programme early by health and demographic characteristics.

Characteristic	OR	95% CI	P value
COPD	0.48	(0.36 to 0.65)	<0.001
Type 2 diabetes	0.75	(0.56 to 0.99)	0.049
Fibromyalgia	0.61	(0.40 to 0.93)	0.020
Prostate cancer	0.18	(0.04 to 0.78)	0.022
Female (vs male)	0.78	(0.71 to 0.86)	<0.001
Age group 1 (youngest vs oldest)	1.36	(1.04 to 1.79)	0.025
Age group 4 (vs oldest)	0.77	(0.61 to 0.98)	0.030

Multinomial logistic regression results comparing participants who left early with those who completed.

COPD, chronic obstructive pulmonary disease.

Table 3. Odds of not participating by health and demographic characteristics.

Characteristic	OR	95% CI	P value
Musculoskeletal conditions	0.31	(0.27 to 0.38)	<0.001
Falls	0.46	(0.36 to 0.58)	<0.001
CABG	0.25	(0.10 to 0.60)	0.002
Angioplasty or stent	0.22	(0.07 to 0.72)	0.012
Stroke or TIA	0.57	(0.40 to 0.81)	0.001
Male (vs female)	0.81	(0.71 to 0.94)	0.005
No weight management referral	0.53	(0.43 to 0.65)	<0.001
Not reporting loneliness	0.58	(0.38 to 0.89)	0.012

Multinomial logistic regression results comparing participants who declined to engage after referral with those who completed.

CABG, coronary artery bypass grafting; TIA, transient ischaemic attack.

Referral sources and outcomes

Within the PARS/MCRactive coding framework, social prescribing referrals (eg, via Be Well) are categorised as ‘outreach’ at the point of referral. Participants referred through outreach were over six times more likely to remain in the ‘Intends to Participate’ category rather than complete the programme (OR=6.08, p<0.001). In contrast, referrals from clinical sources such as GPs, nurses and physiotherapists were associated with significantly higher odds of completion; these referrer types showed reduced likelihood of early dropout relative to completion (all p<0.05) and indicate that clinical referrals carry greater predictive value for engagement than outreach (online supplemental table S3)

Geographical patterns

Online supplemental table S2 shows ward-level variation in participation outcomes (see online supplemental file 2). Several highly deprived wards, such as Harpurhey (OR=0.25, p=0.005 for Intends to Participate vs Completed), Charlestown (p=0.020–0.047 across comparisons) and Moston (p<0.001–0.021) demonstrated significantly lower odds of non-completion relative to completion across multiple outcome categories. In contrast, wards that are generally less deprived, including Deansgate (OR=2.16, p=0.038), Piccadilly (OR=4.20, p=0.006) and Stretford and Humphrey Park (OR=6.69, p=0.020), showed higher odds of non-completion or reduced engagement.

Subgroups identified through ML

Random Forest models accurately predicted who would complete the programme and who would leave early (table 4). Early engagement and deprivation emerged as the strongest predictors (table 5). K-means clustering identified three participant subgroups: ‘Committed Completers’, ‘At-Risk Starters’ and a smaller ‘Administrative’ group (online supplemental table S1).

Table 4. Random Forest model performance.

Scenario	Accuracy	Precision	Recall	F1
Completed vs Left Early	0.83	0.84	0.99	0.91
Completed vs Others	0.86	0.78	0.81	0.80
Left Early vs Others	0.92	0.50	0.05	0.09

Classification accuracy, precision, recall and F1 scores for three binary outcome comparisons.

Table 5. Top five predictors from Random Forest models.

Scenario	First	Second	Third	Fourth	Fifth
Completed vs Left Early	Total attended	IMD score	Ward referral count	Current age	Age at referral
Completed vs Others	Total attended	Total booked	Total absent	IMD score	Session cancelled
Left Early vs Others	Total attended	Total absent	IMD score	Ward referral count	Session cancelled

Most important variables identified for each outcome comparison, ranked by feature importance.

IMD (Index of Multiple Deprivation, 2019), Area-level measure of relative deprivation in England.

Discussion

This study examined factors influencing engagement and adherence within Manchester’s ERS using both regression and ML approaches. Completion rates were lower than national benchmarks; however, high referral volumes from deprived areas indicated strong outreach, reflecting patterns reported in recent evaluations of UK ERS delivery.¹² Age, deprivation and psychosocial referral reasons consistently predicted dropout or non-participation, aligning with evidence showing that socio-demographic and psychosocial need remain central determinants of ERS engagement.¹³ Outreach strongly shaped outcomes, emphasising the importance of programme design and flexible, person-centred delivery in supporting sustained engagement. Finally, exploratory ML analyses identified at-risk subgroups based on early engagement, echoing emerging work proposing that personalised interventions support individuals with complex barriers to participate in ERS.¹⁴

Uptake, dropout and completion

Manchester’s completion rate of 34.6% was lower than the national averages of 43–50%.^{2 4} However, it is important to note that many national reviews are based on considerably smaller samples, often involving hundreds rather than thousands of participants, with most studies reporting sample sizes between 68 and 347 and only a minority exceeding 1000.^{2 15} In contrast, this study draws on a whole-system dataset of 11 909 referrals, including 62.9% of participants from the most deprived IMD quintile. This more comprehensive dataset reduces selection bias and provides a more realistic estimate of adherence in a densely populated, highly deprived metropolitan setting. Several contextual factors help explain the lower completion rate observed. Manchester serves a younger, more diverse and substantially more deprived population than many comparator regions. Deprivation is consistently associated with reduced engagement with health services and a barrier to sustained participation in structured programmes.^{3 7} Individuals living in deprived urban areas often face competing financial, social and occupational demands, alongside poorer baseline health and higher psychosocial stress, all of which limit adherence even when initial intent is strong.^{5 16} In addition, inner-city populations frequently present with more complex multimorbidity, mental health concerns and fluctuating symptoms, which may disrupt routine attendance.⁶ These structural and psychosocial pressures partly explain the lower completion rate observed.

Furthermore, Manchester’s unusually high referral volumes from deprived communities suggest that the service is successfully widening access to groups who are typically underrepresented in ERS participation nationally.⁷ While this is a strength, it also introduces a greater proportion of individuals facing significant barriers to sustained engagement. This combination of broad access and higher underlying vulnerability may therefore depress completion rates relative to schemes in less deprived or less complex populations. For these reasons, direct comparisons with wider regions should be made cautiously, as uptake and adherence are known to vary substantially between dense urban centres and areas with mixed or rural deprivation profiles.^{15 17}

Predictors of adherence and dropout

Regression analyses showed that younger participants, particularly those aged 16–24 years, and those referred for psychosocial reasons, such as long covid, loneliness, anxiety or low mood, were less likely to complete the scheme. This aligns with previous work demonstrating that psychosocial complexity can reduce motivation, confidence and emotional capacity to engage consistently in structured activity.^{5 6} Psychosocial referrals often involve fluctuating symptoms, low mood, social isolation and reduced self-efficacy, all of which are known barriers to sustained behaviour change and adherence.^{18 19} In contrast, participants referred for musculoskeletal or metabolic conditions (eg, diabetes, obesity, weight management) frequently recognise a clearer clinical rationale for participation and therefore engage initially, but may still struggle to maintain attendance in the absence of ongoing support, particularly when symptoms worsen or pain limits activity.^{2 17}

Age and sex differences also reflected broader national evidence. Younger adults consistently show lower adherence to ERS and health programmes,^{3 15} and younger men in deprived communities appear particularly vulnerable to early dropout due to competing life pressures, lower health service engagement and stigma around help-seeking.^{4 7} Together, these factors help explain why these groups differ in their likelihood of completing the programme and highlight the importance of tailored support for individuals whose circumstances make adherence more challenging.

Referral pathways

Referral type was strongly associated with outcomes. In this study, outreach referrals specifically referred to social prescribing pathways, including those made through the Be Well service. These routes were effective at generating intent to participate, but this intent often failed to translate into completion, reflecting the well-documented intention–behaviour gap.¹⁹ One explanation relates to the perceived authority and legitimacy of the referrer. Referrals made by GPs, nurses, and physiotherapists were associated with higher completion rates, likely reflecting greater perceived trust and professional credibility attributed to clinical referrers, which may increase the likelihood of acting on their recommendations.20,22 In contrast, outreach referrals via social prescribing may carry less perceived clinical authority, requiring additional support to sustain engagement.²³ Consequently, outreach referrals may require additional early support, such as structured follow-up, warm handovers and personalised contact, to help convert intention into engagement. These findings suggest that while social prescribing expands reach, enhancing referral quality and early contact processes is essential to improve completion.

Geographical variation

Ward-level analyses demonstrated meaningful variation across Manchester, with some highly deprived wards (eg, Harpurhey, Moston, Charlestown) showing comparatively strong completion among those who engaged. This pattern aligns with the principle of proportionate universalism, which argues that interventions should be available to everyone but delivered with an intensity proportionate to the level of need,¹⁶ such as offering more frequent contact or additional support in communities facing greater disadvantage. In practice, this means that communities facing greater structural disadvantage may experience greater benefit once they are successfully engaged, because the intervention addresses an unmet need that is more pronounced in these areas. The findings, therefore, challenge the assumption that deprivation uniformly predicts poorer outcomes. Instead, they suggest that when access barriers are overcome, and individuals in deprived wards begin the programme, the ERS can be particularly impactful²⁴ despite previous research consistently reporting lower uptake and adherence among individuals living in more deprived areas.^{25 26} This underlines the importance of locally responsive service design, with tailored engagement strategies that recognise both the higher need and the strong potential for benefit in deprived communities.

Machine learning insights

ML added value by modelling complex, non-linear interactions that are not easily captured by regression. Whereas regression assumes linear, additive relationships, the Random Forest models detected combinations of factors, such as age, deprivation and early engagement patterns, distinguishing those likely to complete from those at higher risk of dropping out. Early engagement measures (sessions attended, booked and missed) consistently predicted adherence, reinforcing the evidence that the first few weeks are critical for adherence.²⁷ Clustering analysis also highlighted a distinct high-risk subgroup of younger, deprived males with low early attendance, a pattern consistent with recent ML work identifying similar dropout-risk profiles in community exercise programmes.²⁸ Together, these findings indicate that younger males living in deprived areas, particularly those who do not engage early, require the greatest support at the point of entry into the programme.

These insights offer a practical foundation for targeted intervention. ML-derived ‘risk flags’ could be incorporated into referral management systems to trigger proactive follow-up, helping staff direct additional support to individuals most likely to disengage. Embedding predictive outputs as simple risk indicators, updated at referral and again after early attendance, would provide a structured way to prioritise contact, allocate resources and support, rather than replace practitioner judgement.

Clinical and policy implications

This study highlights several priorities for strengthening exercise referral schemes. Early engagement appears crucial, with rapid follow-up and triage-style contact helping convert referrals into active participation. Improving referral quality through clearer guidance and better integration with clinical pathways may reduce inappropriate or low-commitment referrals. The findings also highlight an ‘intent to engagement gap’ among participants referred through social prescribing: while outreach generates interest, it is less successful in converting intent into sustained attendance. Incorporating brief assessments of readiness to change, clearer communication about programme expectations and more structured ‘warm handover’ processes may help ensure outreach referrals reflect genuine preparedness to participate.

Targeted support is particularly important for younger males and individuals referred for psychosocial reasons, who were consistently more vulnerable to dropout. Closer community integration, such as co-designing outreach approaches with local partners, may further enhance retention in deprived wards. The findings also support integrating simple predictive tools within routine workflows. ML-derived risk indicators, updated at referral and after early attendance, could help staff proactively identify individuals at heightened risk of disengagement and prioritise follow-up, complementing practitioner judgement.

Finally, these implications should be interpreted within the context of Manchester; comparisons with other regions must account for differences in deprivation levels, service models and population profiles.

Limitations

This study has several limitations. Its retrospective and cross-sectional design restricts the ability to draw firm causal conclusions, as associations observed in such datasets may simply reflect underlying patterns rather than directional effects. The use of broad age bands limits the precision with which demographic differences can be understood, and the exclusion of ethnicity, due to high levels of missing data, constrains the extent to which equity-related patterns can be explored, a recognised challenge when working with incomplete public health datasets.²⁹ The machine-learning analyses are exploratory and remain unvalidated outside this sample; it is well established that predictive models in health settings require external testing before informing practice.³⁰ Ward-level comparisons should be interpreted cautiously, given local contextual variation and the reliance on administrative records, rather than participant-reported experience, means that important psychosocial or behavioural factors may not have been fully captured.

Conclusion

Completion rates in exercise referral schemes remain below national benchmarks, but Manchester’s model demonstrates a strong reach into deprived communities. Age, deprivation, psychosocial referral reasons and referral pathways all shaped outcomes, while ML highlighted the added value of early engagement and risk profiling. Strengthening referral quality, supporting at-risk subgroups and embedding predictive tools may improve retention and ensure ERS delivers equitable benefits. These findings should be interpreted in the context of a single metropolitan setting; comparisons with wider regions must consider differences between dense urban centres and areas with more mixed or rural deprivation.

Data availability statement

Data are available upon reasonable request.