Week 5 of 10
Finding the people who need it
Proactive care depends on finding an impactable cohort, and the risk prediction tools most systems reach for are weaker than the confidence placed in them.
The framework asks for a cohort, not a waiting list
The Neighbourhood Health Framework puts population health management inside a specific promise. General practice is to be "empowered to better manage the health of their population by incentivising proactive population health management", and the framework names the method: "risk stratification, long-term condition management, secondary prevention and better continuity of care", directed at "patients with frailty" (Neighbourhood Health Framework, 2026). The neighbourhood health guidelines for 2025/26 are more explicit about the machinery. Systems are told to "apply a single, consistent system-wide population health management method" that will "segment and risk stratify populations, based on complexity and forecasted resource use", and, where no tool exists, to work with the NHS Federated Data Platform team to select one (Neighbourhood Health Guidelines, 2025).
That instruction folds three different questions into one phrase, and this week treats them as separate. Segmentation asks who is in the population and what mix of needs they carry. Risk stratification ranks individuals by the probability of a defined outcome, usually an unplanned admission. Impactability asks a different question again: which of those people would actually benefit from the care on offer. A system can be competent at the first, adequate at the second, and still never have asked the third.
The guidelines show what is at stake in the choice. NHS England analysis from adapted Bridges to Health data puts the cohort of adults, children and young people with complex needs who depend on several services at around 7% of the population, carrying around 46% of hospital costs, before narrowing the early focus to "around 2% to 4% of the population". The framework makes a related calculation for the first integrated neighbourhood teams: people over 75 with frailty, people at end of life and care home residents are 3 to 5% of the population, yet "over 25% of non-elective admissions and 50% of bed days" (Neighbourhood Health Framework, 2026). Both figures describe concentration. Neither says the concentration was caused by care that could have been delivered differently, which is the only version of the claim that supports acting on it.
What a risk score can and cannot tell you
A risk model is a statistical rule built from historic records. It chooses variables, weights them by their association with a future event, and returns a score. Prior admissions and age are usually the strongest terms, so the model is partly a measure of what the health service already did to someone. Validation then reports two different properties. Discrimination, usually the c-statistic or area under the curve, asks whether the people who experience the event tend to score higher than those who do not; 0.5 is no better than chance, and 1.0 would be perfect. Positive predictive value asks the question a clinician asks: of the people the model puts in the top slice, what proportion go on to have the event? The two trade off, and the choice of threshold decides which the system gets.
Wallace and colleagues reviewed validated models for predicting emergency admission in community-dwelling adults, screening 18,983 records and including 27 models, 11 of them developed in the United Kingdom (Wallace et al., 2014, Medical Care). Models built on administrative or clinical records discriminated better than those built on self-report, with c-statistics of 0.63 to 0.83 against 0.61 to 0.74. The 6 models above 0.8 all used prior health care use, multimorbidity or polypharmacy, and diagnoses or prescribed medicines. That is a fair account of the ceiling: a good model built on the data a system already holds lands between 0.7 and 0.8.
England's most widely distributed tool, the Combined Predictive Model, performs at the lower end in practice. The Nuffield Trust built a model on 1.8 million adults and compared it with the Combined Predictive Model (Nuffield Trust, 2013). At the top 1% threshold, the Combined Predictive Model returned a positive predictive value of 0.405 and a sensitivity of 0.060 at the same time. In plain terms, about 40% of the people it flagged were admitted within the year, and it found about 6% of everyone who was admitted. The report's own summary is that although "reasonable accuracy (in terms of positive predictive value) could be obtained, the sensitivity of models was generally low". A reader handed such a model should ask which number matters for the intended use, and expect the answer to be sensitivity: the people the model misses are the ones the service will never contact.
The highest risk is not the most helpable
Geraint Lewis set out the problem in 2010. Predictive models "can be used to identify people at high risk of unplanned hospitalization, although some of the high-risk patients they identify may not be amenable to preventive care" (Lewis, 2010, Milbank Quarterly). Impactibility models were proposed to find the subset for whom preventive care is expected to succeed, by prioritising conditions that respond to upstream care, excluding patients unlikely to respond, or matching the intervention to the person. Two cautions come with the idea. High-cost patients regress towards the mean without any intervention, so some people a model flags would have improved anyway, and Lewis is explicit that impactibility has "important implications for equity and access", because a tool that excludes likely non-responders can move care away from those who need it most.
The idea has not become practice. Orlowski and colleagues ran two workshops with 37 NHS analysts, policy makers, academics and think-tank staff and found the concept "not clearly understood", generally treated as a cross-sectional way of naming patients rather than an iterative judgement about response (Orlowski et al., 2024, BMJ Open). Their account of the underlying problem is the sharpest available. Risk stratification "has not consistently led to improvements in health outcomes across the population"; it cannot distinguish people whose high score is chance from those who need care; and participants warned that a model built on incomplete primary care and social data risks widening the inequality it was meant to reduce. They also name the mechanism that produces invisible need: "less care accessed is mistakenly equated to healthier status by algorithms".
The critique of the risk-first approach is not new. Roland and Abel argued in the BMJ that most attempts to reduce emergency admissions concentrate on people at high risk, that the supporting evidence is weak, and that there are "fundamental flaws in this approach", noting that a widely quoted saving of around 2.3 billion pounds from reducing admissions among frequent attenders rests on an estimate rather than a trial (Roland and Abel, 2012, BMJ).
Set that against frailty, where case finding has a specified cohort and a specified response. Clegg and colleagues built the electronic frailty index from routine primary care records for 931,541 patients, using 36 deficits drawn from 2,171 clinical codes (Clegg et al., 2016, Age and Ageing). Severe frailty carried an adjusted hazard ratio for one-year mortality of 4.52 against fit patients, with "good to moderate discrimination but low calibration estimates". The tool works because what follows it is defined: a structured review, a care plan and a named coordinator, rather than a letter.
The bias that comes from training on existing service users
Every model of this kind is trained on records of people who have already used services, and that one fact carries most of the equity risk. Obermeyer and colleagues examined a widely used commercial algorithm that selected patients for extra care and found that at the same risk score Black patients were considerably sicker than White patients (Obermeyer et al., 2019, Science). The cause was the label. The algorithm predicted health care costs rather than illness, and because less money is spent on Black patients at the same level of need, an accurate cost model produced systematically biased health predictions. Correcting the target would have raised the proportion of Black patients identified for additional help from 17.7% to 46.5%. The same logic applies to any local tool: if the outcome predicted is service use, the model has learned who gets served and will reproduce the pattern while appearing accurate.
Data quality compounds it. Mathur and colleagues profiled ethnicity recording in English primary care and hospital data and found that 27.1% of all patients in the primary care dataset had ethnicity recorded, rising to 78.3% for patients registered since April 2006; in hospital data, 79.4% of inpatients, 46.8% of outpatients and 26.8% of accident and emergency patients had it recorded (Mathur et al., 2014, Journal of Public Health). Combining the two sources raised completeness to 97%, but the authors concluded that the "poor completeness and consistency of A&E and outpatient data render these sources unreliable". A system that runs an equity audit on a field missing for three quarters of its emergency records is auditing the absence, not the need. The guidelines foresee the gap and ask systems to add "wider quantitative and qualitative insight into groups that might be under-represented in NHS datasets, for example, people with severe mental illness or learning disability or autistic people" (Neighbourhood Health Guidelines, 2025).
The groups named there are also the groups whose needs are largest. Barnett and colleagues analysed 1,751,841 patients in 314 Scottish practices and found 23.2% multimorbid, with most people over 65 affected but a larger absolute number under 65; onset arrived 10 to 15 years earlier in the most deprived areas, and physical and mental health conditions coincided in 11.0% of the most deprived group against 5.9% of the least deprived (Barnett et al., 2012, The Lancet). Walker and colleagues pooled 148 studies and put the relative risk of all-cause mortality for people with mental disorders at 2.22 with 67.3% of deaths from natural causes and a median of 10 years of potential life lost (Walker et al., 2015, JAMA Psychiatry). A tool tuned on unplanned admissions will not find these people, because their need has not yet converted into the acute activity the model counts.
The equity audit to run before you act on a model
An equity audit is a blunt instrument, and it runs before a cohort tool changes anyone's caseload. It asks four questions. Who is in the training data, and who is missing from it? Which groups does the chosen threshold exclude, and what happens to them afterwards? What is the model actually predicting, cost or clinical need? Does the model's accuracy hold within each subgroup, or only on average? The literature supports each. Rajkomar and colleagues argue that machine learning systems should encode "principles of distributive justice" in model design, deployment and evaluation, choosing between equality in outcomes, in performance and in resource allocation rather than leaving the choice implicit (Rajkomar et al., 2018, Annals of Internal Medicine). The guidelines already require part of the audit, asking for outcomes to be analysed "by population demographics, deprivation, age, ethnicity, disability" and inclusion health groups (Neighbourhood Health Guidelines, 2025).
None of it is possible without the data sharing the framework also demands. Among the minimum requirements for 2026/27 is that integrated care boards "confirm plans for having the appropriate data-sharing arrangements in place" so that patients can be identified and results evaluated (Neighbourhood Health Framework, 2026). The guidelines specify what that requires: a person-level, linked dataset spanning general practice, community health, mental health, acute care, social care and public health, broadened where possible to housing, employment and education, supported by data sharing and processing agreements and developed with the National Data Guardian (Neighbourhood Health Guidelines, 2025). A cohort tool built on one provider's record will find one provider's patients.
Start from a cohort, or from a predicted risk
This is the week's contested question, and the evidence points one way. The framework asks systems to plan on the basis of "patient risk register analysis", which is an argument for the model; but it also names its cohorts in advance and asks for a care model for each: frailty, care home residents, housebound patients, end of life, cardiovascular disease, diabetes, chronic obstructive pulmonary disease, dementia, mental health conditions, and children and young people (Neighbourhood Health Framework, 2026). The 10 Year Health Plan goes further, describing a service that "predicts and prevents ill health" and a genomics population health service built on predictive analytics, while conceding that "it does little to know about an individual's risk of disease if that knowledge does not lead to behaviour change, or different approaches to care" (10 Year Health Plan, 2025).
The defensible position is narrower than either side usually states. Start from a cohort whose needs are known and whose care model is already specified, use risk stratification to find the people inside it, and audit the model for who it misses. That is what the frailty index allows and what a general-purpose admission model does not. A c-statistic of 0.7 tells a commissioner nothing about whether the care on offer will help; it says only that the ranking beats chance. The test a neighbourhood team should apply is not the model's discrimination but the audit behind it: which groups are under-recorded, who is excluded by the threshold, and what the service will do for the people it does not select.
The contested question
Whether population health management should start from predicted risk at all, or from a defined cohort whose needs are known and whose care model is already specified.
Working across professions
Whose data, whose risk
The cohort a team chooses is a professional judgement wearing a statistical coat. Give the group one dataset, ask three professions to name the cohort that matters most, and then examine why they differ. The exercise is about defending a cohort choice to a colleague who would have chosen differently.
For your profession
Written for the groups City St George's trains. Read your own first, then the two nearest to you: the disagreements between them are the working material.
General practice and primary care (including physician associates)
Physician Associate Studies (Pre-Registration) MPAS
What changes for you. The practice works from a risk register rather than a diary. The framework's 2026/27 requirement is an initial plan to reduce non-elective admissions and bed days based on patient risk register analysis, so somebody in the practice must own the cohort list, decide who is contacted, and record what happened. The guidelines also ask for a person-level linked dataset, which changes what the practice shares beyond its own clinical system.
The decision on your desk. Which of the flagged patients the practice contacts first, and whether a physician associate or the GP runs the structured review. The score produces a ranking, and the practice decides who is actually seen.
What goes wrong if this is done badly. The ranked list is treated as a clinical judgement. The Combined Predictive Model found about 6% of the people who were admitted, so a caseload built on the top slice misses most future admissions and fills with patients whose high score came from previous activity.
Community nursing and district nursing
Public Health and Specialist Community Nursing MSc/PGDip
What changes for you. District nursing receives the caseload the model selects, with a named coordinator, a structured review and a care plan for each person the framework lists as frailty, housebound or end of life. The electronic frailty index can be built from routine records for this group, but its calibration is low, so nurses will meet patients whose measured frailty overstates their actual need.
The decision on your desk. Whether to accept a caseload selected by a risk score or to add patients the district nurse already knows are deteriorating, and how that professional disagreement is recorded in the shared record.
What goes wrong if this is done badly. Unplanned visits rise without matching capacity. The guidelines put the complex needs cohort at around 7% of the population and around 46% of hospital costs, so a caseload drawn from that group without extra staff turns planned work into reactive cover.
Mental health nursing and mental health practice
Mental Health Nursing BSc (Hons)
What changes for you. People with co-occurring severe mental illness are largely invisible to an admission risk model, because their need has not yet become acute activity. The guidelines ask systems to add qualitative insight about groups under-represented in NHS datasets, naming people with severe mental illness, and the framework expects their care to stay within community mental health services under a joined-up plan. Mental health nursing supplies that insight and holds the physical health checks.
The decision on your desk. Whether the mental health practitioner works from the modelled cohort list or keeps a separate caseload, and what evidence is entered when a patient the score misses is already known to be at risk.
What goes wrong if this is done badly. The cohort is defined by what the data can see. With a pooled relative risk of death of 2.22 and 67.3% of deaths from natural causes, leaving this group outside the neighbourhood cohort removes the patients whose deaths are most preventable.
Physiotherapy
Physiotherapy BSc (Hons)
What changes for you. Physiotherapy enters after the cohort has been chosen, because a score for unplanned admission does not identify who will regain function. Severe frailty carries an adjusted hazard ratio for one-year mortality of 4.52, and the framework expects the frailty cohort to receive assessment, care planning and follow-on support. The practical change is a standing review slot for the known frailty list rather than referral after a fall.
The decision on your desk. Which patients in the cohort are seen before their first admission rather than after it, and whether first contact practitioner time goes to new referrals or to the frailty list.
What goes wrong if this is done badly. Rehabilitation is commissioned as outpatient work with a waiting list. The score keeps producing names while the window in which falls risk and function can be changed closes, and the admissions the cohort was meant to prevent still happen.
Occupational therapy
Occupational Therapy BSc (Hons)
What changes for you. Occupational therapy is where a cohort decision becomes a change in daily life. The data show multimorbidity arriving 10 to 15 years earlier in the most deprived areas and physical and mental conditions coinciding in 11.0% of that group against 5.9% of the least deprived, so housing, equipment and adaptation work clusters in the same patients. The change is a named route into housing and social care before the score has reason to flag anyone.
The decision on your desk. Whether to act on the equipment and adaptation needs the equity audit reveals, or to wait until a patient appears on the admission risk list, and who authorises the housing referral.
What goes wrong if this is done badly. Adaptations arrive after crisis. Measured frailty drives the list while daily function does not, so equipment and reablement reach people who have already lost independence and the pathway stalls between health and the council.
Adult social care
Partner profession: local authority adult social care
What changes for you. Adult social care is a partner in this work, not a provider the tool selects for. The framework makes local authorities joint owners of the neighbourhood plan and of the data-sharing arrangements, and the linked dataset it asks for spans social care records alongside general practice, community health, mental health and acute care. The practical change is a seat at the table where the cohort and the threshold are set.
The decision on your desk. Whether the council contributes its social care records to the linked dataset on the terms the health partners propose, and what safeguarding and housing data it withholds from that dataset.
What goes wrong if this is done badly. The cohort is drawn from health records alone. Patients whose need appears first as a housing, safeguarding or care need are absent from the model, and the council inherits the crisis admission the neighbourhood service was meant to prevent.
Public health and local government
Public Health MPH
What changes for you. Public health holds the analysis the equity audit depends on. The guidelines ask for outcomes analysed by demographics, deprivation, age, ethnicity, disability and inclusion health groups, and for insight about groups under-represented in NHS datasets, while one English study found ethnicity recorded for only 27.1% of primary care patients and 26.8% of accident and emergency patients. The audit becomes a routine product rather than a one-off study.
The decision on your desk. Whether the system publishes the equity audit before the tool is switched on, and who signs off the threshold at which each cohort is cut.
What goes wrong if this is done badly. The tool goes live without the audit. A cost-based score reads lower spending as better health, the deprivation gradient the council already measures widens, and the first evidence of harm arrives as a difference in admissions rather than as a warning.
Psychological therapies and applied psychology
Professional Doctorate in Counselling Psychology DPsych
What changes for you. Psychological therapy is largely absent from a model built on admissions and diagnoses. Mental disorders carry a pooled relative risk of all-cause mortality of 2.22, so the people this service sees are high in need and low in visibility, and the guidelines ask for insight about exactly such under-represented groups. The change is that psychological formulation becomes part of why a patient is on the cohort list at all.
The decision on your desk. Whether people who do not engage with a care offer are removed from the cohort or kept on it, since the impactability literature warns that excluding likely non-responders can move care away from those who need it most.
What goes wrong if this is done badly. Non-engagement is read as low risk. Less contact with services is mistaken for better health by the algorithm, and patients who attend repeatedly without a recorded physical cause are routed back to the GP rather than into therapy.
Speech and language therapy
Speech and Language Therapy BSc (Hons) and MSc
What changes for you. Risk models trained on cost and previous admissions cannot see the patients speech and language therapy protects, because that need produces neither. Dysphagia is a clinical state rather than a billing line, and communication difficulty rarely generates a recorded admission until aspiration or a crisis already has. Obermeyer's cost-based label is the general case.
The decision on your desk. Whether to add a clinical route into the cohort for people whose swallowing or communication need carries aspiration risk but no admission history, or to let the risk score decide who the team contacts first.
What goes wrong if this is done badly. If the cohort is the model's output, patients with unrecorded dysphagia are contacted only after an aspiration admission, communication need never enters the list, and the patients therapy would have kept out of hospital become the ones it cannot reach in time.
Critical commentary
Where this week is argued with
Full-text commentaries and critical analyses from peer-reviewed journals. The policy column tests the claim against what government actually commissioned. The practice column tests it against what staff can deliver. Read one from each before the contested question.
Policy-facing
Commentary from the PROPHET consortium arguing that robust evidence and clinical utility are not enough to put risk-stratified prevention into practice, because governance, reimbursement, workforce and data infrastructure decide adoption. It warns that stratification can widen inequalities and medicalise the well, so the binding constraint is delivery rather than prediction. It supports this week's scepticism about tool-led case finding.
Commentary noting that prevention policy has made high-risk identification a cornerstone while dementia algorithms rest on unrepresentative cohorts, inconsistent outcome definitions, internal validation and undisclosed cut-offs. It argues prediction is limited less by modelling technique than by data and evaluation, so apparent performance will not survive translation. It strengthens this week's thesis and locates the weakness in evidence, not maths.
Editorial arguing that the English neighbourhood health programme has no funding vehicle for work spanning health and social care, so community health and wellbeing worker schemes close before replacement money arrives. It backs those workers' universal household visiting and puts the constraint in fragmented budgets rather than model performance. It resists the week's focus on better identification alone.
Practice-facing
Externally peer reviewed commentary proposing four iterative steps: empanel patients sharing a risk, match an intervention to each panel, surveil care gaps, then evaluate effect. It states that a risk stratification algorithm is not flawless and needs clinician review, and that multimorbidity, siloed data and payment models block adoption. It supports the thesis while shifting attention from prediction to intervention.
Commentary from the team that validated CONECT-6, reporting that six yes-or-no questions about complex need had to be reworded and supported in use. Validation reports sensitivity of 90 per cent but a positive predictive value of 49 per cent, so half of those flagged are false positives. It supports the week by putting the weakness in administration, not detection.
Editorial reviewing two studies in which machine learning outperformed the National Early Warning Score, and arguing that risk tools should augment rather than replace clinical judgement. It pushes back on this week's scepticism, yet concedes that single-site data, opacity, unresolved fairness and unproven clinical benefit still hold adoption back. It lands on the optimistic side.
Key sources
Study drawing on interviews with representatives of thirty American organisations that build, use or assess predictive models, proposing that impactibility models identify the subset of at-risk patients for whom preventive care is expected to succeed. It sets out three modelling approaches and notes that excluding likely non-responders has consequences for equity and access. It is the origin of this week's distinction between predicting risk and predicting benefit.
Analysis of a widely used commercial algorithm that enrolled patients in extra care, showing that at the same risk score Black patients were considerably sicker than White patients. The bias arose because the algorithm predicted health care costs rather than illness, and correcting the target would have raised the share of Black patients identified for additional help from 17.7% to 46.5%. It is this week's central evidence on the bias that comes from training on existing service users.
Systematic review that screened 18,983 records and included 27 validated models predicting emergency admission in community-dwelling adults, eleven of them developed in the United Kingdom. Models built on administrative or clinical records achieved c-statistics of 0.63 to 0.83 against 0.61 to 0.74 for models using self-report data, and the six models above 0.8 all included prior health care use. It supplies the performance range a reader should expect from a real tool.
Development and validation study of an electronic frailty index built from routine primary care records for 931,541 patients, using 36 deficits drawn from 2,171 clinical codes. One-year adjusted hazard ratios for mortality rose from 1.92 for mild frailty to 4.52 for severe frailty, with good to moderate discrimination but low calibration. It shows what case finding looks like when the cohort and the following intervention are both defined.
Cross-sectional analysis of 1,751,841 people registered with 314 Scottish general practices, finding that 23.2% were multimorbid and that the absolute number with multimorbidity was higher below the age of 65. Onset arrived 10 to 15 years earlier in the most deprived areas, and physical and mental health conditions coincided in 11.0% of the most deprived group against 5.9% of the least deprived. It undercuts any cohort definition built on age alone.
Systematic review and meta-analysis of 203 cohort studies across 29 countries, pooling the relative risk of all-cause mortality for people with mental disorders at 2.22. Two thirds of deaths, 67.3%, were from natural causes, and the median years of potential life lost was 10. It establishes the size of a need that a risk tool based on acute activity is least likely to detect.
Roland M, Abel G (2012). Reducing emergency admissions: are we on the right track? BMJ, 345, e6017.
LANDING PAGEAnalysis article arguing that most attempts to reduce emergency admissions concentrate on people at high risk, that the supporting evidence for doing so is weak, and that the approach contains fundamental flaws. It notes that a widely quoted saving of around 2.3 billion pounds from reducing admissions among frequent attenders rests on an estimate rather than on trial evidence. It frames this week's contested question from the sceptical side.
Qualitative assessment of two workshops attended by 37 NHS analysts, policy makers, academics and think-tank staff, examining how impactibility analysis is understood. Participants treated the concept as unclear, raised incomplete primary care and social data and the risk of widening inequality, and the authors record that risk stratification has not consistently improved population health outcomes. It is the most recent statement of where impactability stands in English practice.
Comparative study of ethnicity recording in the Clinical Practice Research Datalink and Hospital Episode Statistics. Ethnicity was recorded for 27.1% of all primary care patients, rising to 78.3% for those registered since April 2006, and for 79.4% of inpatients but only 46.8% of outpatients and 26.8% of accident and emergency patients, which the authors call unreliable. It shows the data limits an equity audit must state.
Methodological article describing how model design, biases in training data and the interaction of predictions with clinicians can widen health care disparities, and arguing that distributive justice should be built into design, deployment and evaluation. It sets out technical options that make models equal in patient outcomes, in performance or in resource allocation, and when each should be preferred. It is the reference for the equity audit this week asks the reader to design.
Report describing a new model of unplanned admission built on 1.8 million adults and comparing it with the Combined Predictive Model used across England. At the top 1% threshold the Combined Predictive Model achieved a positive predictive value of 0.405 with a sensitivity of 0.060, and the authors note that the sensitivity of most United Kingdom models is generally low. It supplies the published performance figures for the tools English systems actually used.
Department of Health and Social Care and NHS England (2026). Neighbourhood Health Framework. GOV.UK.
FULL TEXTThe framework that commits general practice to proactive population health management through risk stratification, names the priority cohorts from frailty and end of life to diabetes and mental health conditions, and sets the 3 to 5% of population against over 25% of non-elective admissions and 50% of bed days. It also makes confirming data-sharing arrangements for patient identification and evaluation a minimum requirement for 2026/27. It is the primary policy object of this week.
Guidance setting out the six core components of a neighbourhood health service, including population health management. It requires systems to apply a single population health management method to segment and risk stratify populations by complexity and forecasted resource use, puts the complex needs cohort at around 7% of the population and around 46% of hospital costs, and asks for outcomes to be analysed by deprivation, ethnicity and disability. It also describes the linked dataset and data-sharing agreements the work needs.
The plan that frames the neighbourhood health service as predictive and preventative, promising a service that predicts and prevents ill health and a genomics population health service built on predictive analytics. It concedes that knowing about an individual's risk of disease achieves little if that knowledge does not lead to behaviour change or a different approach to care. It is the strongest statement of the risk-first position this week tests.