Abstract
BACKGROUND: Falls among hospitalized patients are a critical issue that often leads to prolonged hospital stays and increased health care costs. Traditional fall risk assessments typically rely on standardized scoring systems; however, these may fail to capture the complex and multifactorial nature of fall risk factors. OBJECTIVE: This retrospective observational multicenter study aimed to develop and validate a machine learning-based model to predict in-hospital falls and to evaluate its performance in terms of discrimination and calibration. METHODS: We analyzed the data of 83,917 inpatients aged 65 years and older with a hospital stay of at least 3 days. Using Diagnosis Procedure Combination data and laboratory results, we extracted demographic, clinical, functional, and pharmacological variables. Following the selection of 30 key features, 4 predictive models were constructed: logistic regression, extreme gradient boosting, light gradient boosting machine (LGBM), and categorical boosting (CatBoost). The synthetic minority oversampling technique and isotonic regression calibration were applied to improve the prediction quality and address class imbalance. RESULTS: Falls occurred in 2173 (2.6%) patients. CatBoost achieved the highest F(1)-score (0.189, 95% CI 0.162-0.215) and area under the precision-recall curve (0.112, 95% CI 0.091-0.136), whereas LGBM had the best calibration slope (0.964, 95% CI 0.858-1.070) with good discrimination (F(1)-score 0.182, 95% CI 0.156-0.209; area under the precision-recall curve 0.094, 95% CI 0.078-0.113). Logistic regression had the lowest discrimination (F(1)-score 0.120, 95% CI 0.100-0.143). Shapley Additive Explanations analysis consistently identified low albumin, impaired transfer ability, and the use of sedative-hypnotics or diabetes medications as major contributors to fall risk. In incident report analysis (n=435), 49.2% of falls were toileting-related, peaking between 4 and 6 AM, with bedside falls predominating in high or very high risk groups. CONCLUSIONS: CatBoost and LGBM offer clinically valuable prediction performance, with CatBoost favored for high-risk patient identification and LGBM for probability-based intervention thresholds. Integrating such models into electronic health records could enable real-time risk scoring and trigger targeted interventions (eg, toileting assistance and mobility support). Future work should incorporate dynamic, time-varying patient data to improve real-time risk prediction.