Prediction Models for Early Post-Discharge Falls Among Older Adults Using Machine Learning: A Prospective Cohort Study

Abstract Purpose To construct a prediction model for early post-discharge falls among older adults in Japan using machine learning, leveraging patient information collected during hospitalization. Methods This prospective cohort study was conducted at an acute care hospital in Osaka, Japan. Participants were inpatients aged ≥ 65 years admitted to the geriatric ward between February 2022 and July 2023. At admission and discharge, 83 patient information items were collected from electronic medical records. The outcome, recorded within three months of discharge, was collected telephonically. Variables used in the model were selected based on statistical analyses and clinical findings, and the model was constructed using five algorithms. Results The analysis included 156 patients [mean age: 78.1 ± 5.8 years; women: 79 (50.6%)], 19 (12.2%) of whom had fallen within three months of discharge. Six variables were used in the model: “alanine aminotransferase (ALT),” “self-assessment of health status,” “decrease in grip strength,” “Clinical Frailty Scale ≥ 4,” “fecal incontinence,” and “urinary incontinence.” The best area under the precision-recall curve (AUPRC) and area under the receiver operating characteristic curve (AUROC) values were obtained using CatBoost (AUROC: 0.801; AUPRC: 0.392; sensitivity: 0.375; specificity: 0.964). The highest sensitivity was found for ExtraTrees (AUROC: 0.719; AUPRC: 0.331; sensitivity: 0.500: specificity: 0.964). Conclusion The results demonstrate the potential of using machine learning to construct a model for predicting early post-discharge falls in hospitalized older adult patients.