Abstract
Accurate prediction of inpatient fall risk is crucial for preventing injuries and improving patient safety in hospitals. The widely used Hester-Davis score (HD) lacks precision, highlighting the need for more advanced models. This retrospective study analyzed 46,695 patients from 17 hospitals across four U.S. states, admitted between January 2018 and July 2022, including 4245 fallers. Four dynamic machine learning models were developed using HD variables alone and in combination with socio-demographics, comorbidities, physiological measures, medications, and timeseries data updated at 8- and 24 h intervals. Among the models, Extreme Gradient Boosting model outperformed HD, achieving an AUC of 0.87 (95% CI 0.86-0.88), compared to HD AUCs of 0.57 (95% CI 0.56-0.58) and 0.62 (95% CI 0.59-0.61) at thresholds of 7 and 20. Key predictors included pre-existing neurological conditions, behavioral abnormalities, oxygen saturation, heart rate, and IV furosemide use. Prospective validation is required for real-time implementation in clinical practice.