Extending eFall Risk Prediction to Working-Age Adults within Mental Health and Learning Disability Services:A Clinical Validation Study

Background: Most fall risk tools are developed for older adults; evidence in working-age populations receiving mental health services is limited.<br/><br/>Objective: To externally validate the existing eFalls prediction model in working-age adults (18–65 years) within a UK integrated teaching National Health Service (NHS) foundation trust and assess the impact of simple recalibration.<br/><br/>Methods: We conducted a retrospective validation using routinely collected electronic health records from a large integrated teaching NHS foundation trust in the north of England. The published eFalls coefficients were applied to derive 12-month fall/fracture risk. Performance was evaluated using discrimination (C-statistic), calibration-in-the-large (CITL), calibration slope, observed-to-expected (O/E) ratio, calibration plots, and decision curve analysis (DCA). A logistic recalibration (intercept and slope) using the original linear predictor was then fitted on the full cohort and applied uniformly to subgroups (sex; mental health, learning disability).<br/><br/>Results: Among 32,410 adults (fall rate 2.07%), the model showed good discrimination (C-statistic = 0.777). Before recalibration, calibration was suboptimal (CITL = 1.36; O/E = 1.60; slope = 1.22). Recalibration restored alignment (CITL ≈ 0; O/E ≈ 1; slope ≈ 1) without changing discrimination. Subgroup analyses revealed degraded performance in learning disability groups (e.g., AUC 0.696–0.739; marked underprediction), whereas sex and mental-health-only groups were closer to overall performance. DCA indicated positive net benefit across clinically relevant thresholds (10–25%).<br/><br/>Conclusions: The eFalls model showed reasonable performance in working-age adults receiving mental health or learning disability services following simple recalibration. However, discrimination was lower among individuals with learning disabilities, suggesting that recalibration alone may be insufficient and that further model refinement and validation in this subgroup are warranted.<br/>