Explainable SHAP-space fall risk profiling from electronic health records for inpatient prevention planning

Abstract Falls in hospital settings are common and costly. They happen for many different reasons, which limits the effectiveness of one-size-fits-all prevention strategies. In a retrospective observational study using electronic health record data from a large German university hospital between 2016 and 2022, we trained an ensemble model to predict inpatient falls. We then derived patient-level SHapley Additive exPlanations (SHAP) and clustered model-detected fallers to identify recurrent SHAP-based risk profiles. The resulting centroids were applied to alerted patients in the held-out test set. The final model achieved an area under the receiver operating characteristic curve of 0.95 and an area under the precision-recall curve of 0.36 on the test set, outperforming a baseline model that used only guideline features. We benchmarked SHAP-space clustering against raw feature-space clustering techniques in this alerted test cohort, finding that SHAP-space K-means showed the largest separation in observed fall incidence. The nine clusters identified were then mapped to clinically interpretable risk profiles. These profiles link model alerts to guideline-consistent prevention domains. However, further external validation studies are needed before clinical deployment.