An interpretable data-driven approach to optimizing clinical fall risk assessment

In this study, we aim to better align fall risk prediction from the Johns Hopkins Fall Risk Assessment Tool (JHFRAT) with clinical expert fall risk perception via a data-driven modelling approach. We conducted a retrospective cohort analysis of 54,209 inpatient admissions from three Johns Hopkins Health System hospitals between March 2022 and October 2023. In the absence of a true fall risk ground truth, we apply proxy labels based on clinician choices for the application of targeted preventative interventions, resulting in a total of 20,208 high-risk encounters and 13,941 low-risk encounters. We employed constrained score optimization (CSO) models to recalibrate the JHFRAT scoring weights, while preserving its additive structure and clinical thresholds. Recalibration refers to adjusting item weights so that the resulting score can order encounters more consistently by the study's risk labels, and without changing the tool's form factor or deployment workflow. The CSO model demonstrated significant improvements over the current JHFRAT in classification alignment with the proxy labels (CSO AUC-ROC = 0.91, JHFRAT AUC-ROC = 0.86). This model performance translates to a weekly average of an additional 35 Johns Hopkins Health System patients who are perceived as high risk (per our proxy labels) being classified by JHFRAT as high risk. The ablation analyses also suggest that the CSO model, though outperformed in prediction metrics by the benchmark black-box XGBoost model, is more robust than XGBoost to variations in risk labeling. Our evidence-based approach provides a robust foundation for understanding risk factor contributions to various indicators of clinician-perceived fall risk. Future research can build upon this foundation to improve risk assessment utility as a decision-support tool in clinical practice.