Development and validation of a machine learning-based prediction model for inpatient fall risk: an informatics solution for empowering precision nursing decision-making

Inpatient falls remain a major challenge in clinical nursing. Traditional tools such as the Morse Fall Scale rely on subjective, static assessments that often fail to capture dynamically changing risks. This study aimed to develop and rigorously validate a machine learning model using electronic health record data to provide an objective, dynamic decision-support tool for fall prevention. We conducted a retrospective cohort study of 45,782 adult inpatients hospitalized between January 2020 and December 2023. A total of 25 clinically relevant predictors were extracted from admission records and daily nursing assessments. Missing data were handled using random forest imputation. Class imbalance (fall rate 0.9%) was addressed by applying the Synthetic Minority Over-sampling Technique exclusively to training folds during ten-fold cross-validation. Five algorithms—logistic regression, random forest, support vector machine, XGBoost, and LightGBM—were compared. Temporal validation was performed using an independent dataset from January to March 2024. Model performance was evaluated using the area under the receiver operating characteristic curve, calibration metrics, decision-curve analysis, and comparison with the Morse Fall Scale where contemporaneous scores were available. SHapley Additive exPlanations values were used to interpret the model. The XGBoost algorithm achieved the best performance, with an area under the curve of 0.92 (95% confidence interval 0.90–0.94) on the internal test set and 0.89 (95% confidence interval 0.86–0.92) on temporal validation. The calibration slope was 0.96 with an intercept of − 0.12, indicating excellent agreement between predicted and observed risks. At the threshold corresponding to 85% sensitivity, precision was 0.24. Decision-curve analysis demonstrated net benefit across all clinically reasonable risk thresholds. The model significantly outperformed the Morse Fall Scale, with an area under the curve difference of + 0.14 ( p < 0.001). The key predictive factors were history of falls, advanced age, and use of sedative-hypnotic medications. The XGBoost model demonstrates strong discrimination, good calibration, and added value over current clinical practice. It represents a rigorously validated prediction tool, though prospective implementation studies are needed before clinical deployment.