Data-Driven Identification of High-Risk Patient Groups for Falls Using SHAP-Space Profiling

Hospital falls are common and costly, yet fall risk is heterogeneous, limiting one-size-fits-all prevention. We aimed to derive interpretable, deployable fall-risk profiles from electronic health records (EHRs) to support subgroup-specific interventions. In a retrospective observational study using EHR data from a large German university hospital (2016–2022; N = 932,102), we trained an ensemble model to predict inpatient falls and computed patient-level Shapley Additive Explanations (SHAP). We then clustered fall-risk-flagged patients in SHAP space to obtain clinically interpretable risk profiles, benchmarked this approach against alternative clustering strategies, and assessed threshold trade-offs for high-risk classification. The ensemble achieved excellent discrimination (AUROC 0.95) and strong performance for the rare outcome (AUPRC 0.36) on a held-out test set. The model outperforms a guideline-features-only baseline and is competitive with prior falls prediction models in the literature. SHAP-space clustering yielded nine distinct, interpretable risk groups with specific risk factors. This allows actionable mapping to guideline domains and enabling tailored interventions based on established risk profiles. Overall, we provide an approach of how explainable, outcome-guided clustering can translate EHR-based fall prediction into tailored interventions in inpatient care.