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

Abstract Background Artificial intelligence-enabled tools are increasingly used to support mental health diagnosis, triage and clinical decision-making. However, limited evidence exists on whether these tools incorporate equity-relevant population characteristics, particularly among people living with multiple long-term conditions (MLTC), who often experience complex health needs and structural inequalities. Methods We conducted a cross-sectional analysis using linked Clinical Practice Research Datalink Gold and Aurum primary care data for 7,186,890 adults aged ≥ 18 years with MLTC in England between 1987 and 2020. Five artificial intelligence-based mental health diagnostic or triage tools identified through a scoping review were operationalised using routinely collected primary care variables. Many symptom-level psychiatric inputs central to the original models could not be captured in routine primary care data and were excluded. Reported tool inputs were mapped to sociodemographic, clinical and social-care factors, including deprivation, ethnicity, age, sex, disability, place of residence and social care involvement, informed by the CORE20PLUS5 and PROGRESS-PLUS frameworks. We calculated an Equity-Variable Coverage Index to quantify the proportion of equity-relevant variables each tool includes as inputs, a measure of input coverage rather than demonstrated fairness. To illustrate the empirical importance of these characteristics, we fitted multivariable logistic regression models for hospital admission, a proxy outcome influenced by system-level and clinical factors, with leave-one-out removal of each equity-related variable. Results Equity domain coverage varied substantially across tools, ranging from 13.0% to 56.5%. Clinical characteristics were more consistently represented than sociodemographic or social-care factors. Drug and alcohol misuse was included in all five tools, whereas age, ethnicity, deprivation, disability status and social care involvement were inconsistently incorporated. Removing ethnicity and age produced the largest reductions in model explanatory performance, followed by outpatient attendance, Accident and Emergency attendance, and drug and alcohol misuse, indicating that several under-represented equity-relevant variables were associated with hospital admission among adults with MLTC. Conclusions Current artificial intelligence-enabled mental health tools show substantial and variable gaps in their coverage of equity-relevant population characteristics available in routine UK primary care data. Embedding equity considerations into tool design and routinely evaluating model inputs and performance using population-scale data will support fair and responsible deployment of mental health artificial intelligence.