The study, published in BMC Psychiatry, was led by Aman Jat and Hajira Dambha-Miller of the Primary Care Research Centre at the University of Southampton, together with colleagues at the University of Manchester and Newcastle University. The team set out to map the equity-relevant inputs of AI-based mental health diagnostic and triage tools onto routinely collected primary care data, using one of the largest and most representative primary care datasets in the world. Their focus was a group that is often invisible in technology design: adults living with multiple long-term conditions, or MLTC, whose complex health needs and exposure to structural inequalities make them especially vulnerable to biased algorithms.

The scale of the analysis is striking. The researchers drew on linked data from the Clinical Practice Research Datalight — specifically its Gold and Aurum databases — covering 7,186,890 adults aged 18 and over with multiple long-term conditions in England, with records spanning 1987 to 2020. This is not a sample or a convenience cohort; it is a population-based picture of primary care as it is actually delivered. The team then identified five AI-based mental health diagnostic or triage tools through a scoping review and systematically mapped the variables those tools require as inputs onto the sociodemographic, clinical and social-care factors that can be captured in routine primary care records.

To structure that mapping, the researchers turned to two established frameworks for thinking about health inequalities. The first is CORE20PLUS5, NHS England’s approach that identifies the most deprived 20 percent of the population as a priority group alongside five clinical focus areas. The second is PROGRESS-Plus, a widely used framework that enumerates the dimensions along which health opportunities are unequally distributed: place of residence, race, ethnicity, culture and language, occupation, gender and sex, religion, education, socioeconomic status and social capital, plus additional stratifying factors such as disability and social care involvement. Using these lenses, the team built an Equity-Variable Coverage Index — a quantitative measure of the proportion of equity-relevant variables that each tool actually includes among its inputs.

The results reveal a landscape of uneven and often patchy coverage. Across the five tools, equity domain coverage ranged from just 13.0 percent to 56.5 percent — meaning that even the best-performing tool incorporated barely more than half of the equity-relevant characteristics the researchers considered, while the worst captured barely one in eight. Clinical characteristics fared considerably better than sociodemographic or social-care factors. Drug and alcohol misuse was the only equity-relevant variable included in all five tools. By contrast, age, ethnicity, deprivation, disability status and social care involvement were incorporated only inconsistently — a finding with real consequences, because these are precisely the factors that shape who becomes ill, who seeks help, and how their care unfolds.

There is an important technical caveat here, and the authors are careful to make it. Many of the symptom-level psychiatric inputs that are central to the original AI models — the granular clinical features that machine learning systems often rely on for mental health diagnosis — could not be captured in routine primary care data and were therefore excluded from the mapping exercise. The Equity-Variable Coverage Index, the researchers stress, measures input coverage rather than demonstrated fairness. A tool that includes more equity-relevant variables is not automatically a fairer tool; but a tool that is blind to deprivation, ethnicity or disability cannot even be assessed for fairness across those dimensions, let alone corrected.

To demonstrate why these variables matter empirically, the team fitted multivariable logistic regression models for hospital admission — used as a proxy outcome influenced by both system-level and clinical factors — and then systematically removed each equity-related variable in turn, a leave-one-out design that reveals how much explanatory work each variable performs. The results were unambiguous. Removing ethnicity and age produced the largest reductions in the model’s explanatory performance, followed by outpatient attendance, Accident and Emergency attendance, and drug and alcohol misuse. In other words, several of the very variables that AI mental health tools most often omit are among the strongest statistical signals associated with hospital admission among adults with multiple long-term conditions.

The implications extend well beyond the United Kingdom. Multimorbidity is one of the fastest-growing challenges facing health systems worldwide, and people with multiple long-term conditions are disproportionately likely to experience mental ill health, poverty and fragmented care. If AI triage tools are deployed in such populations without adequate representation of deprivation, ethnicity, disability or social care involvement, they risk systematically underestimating risk in the groups who need the most support — quietly amplifying the very inequalities that health policy aims to reduce. An algorithm that performs well on average can still fail badly for specific subgroups, and input coverage is the first, necessary condition for detecting such failures.

The study also highlights a structural tension in health AI development. Research-grade models are often trained on rich, purpose-collected datasets containing detailed symptom assessments, validated questionnaires and self-reported social information. Routine primary care records, by contrast, are collected for clinical and administrative purposes, and their coverage of social determinants — housing, employment, social support, care involvement — is notoriously incomplete. When such models are operationalised in real health systems, the gap between what they were designed to consume and what the data infrastructure can supply becomes a source of silent degradation. The authors argue that embedding equity considerations into tool design from the outset, and routinely evaluating both model inputs and performance using population-scale data, is essential for fair and responsible deployment.

Funded by the National Institute for Health and Care Research’s cross-NIHR collaboration on multiple long-term conditions, the study offers a practical template that other health systems could adopt: identify the equity-relevant variables available in routine data, quantify how much of that landscape each AI tool actually sees, and test empirically how much predictive information those variables carry. As mental health AI moves from pilot projects to frontline deployment, this research delivers a clear message to developers, regulators and clinicians alike. Fairness in medical AI does not begin with clever debiasing algorithms; it begins with making sure the model knows who the patient is. On that measure, the current generation of AI mental health tools still has a long way to go.

Subject of Research: Equity-relevant input coverage of AI mental health tools mapped to routine primary care data in adults with multiple long-term conditions

Article Title: Mapping equity-relevant inputs of AI mental health tools to routine primary care data: a population-based study of 7.19 million adults with multiple long-term conditions in England

Article References: Mapping equity-relevant inputs of AI mental health tools to routine primary care data: a population-based study of 7.19 million adults with multiple long-term conditions in England. (n.d.). https://doi.org/10.1186/s12888-026-08723-5

Image Credits: AI Generated

DOI: 10.1186/s12888-026-08723-5

Keywords: artificial intelligence, mental health, health equity, multimorbidity, electronic health records, primary care, CPRD, health inequalities, machine learning, clinical decision support, social determinants of health, NHS