Risk Prediction Models for Falls Among Older Adults Inpatients with Cognitive frailty: Machine Learning Study Based on Comprehensive Geriatric Assessment

Background Falls are a major cause of disability in older adults, and cognitive frailty confers greater risk than isolated deficits. However, prediction models seldom target this subgroup. This study aimed to develop machine learning (ML)-based fall risk models for cognitively frail older adults using using Comprehensive Geriatric Assessment (CGA) data. Methods We included 814 hospitalized older adults with cognitive frailty, and corrected class imbalance using random under-sampling (n = 332). Eleven machine learning (ML) algorithms were trained using two feature selection strategies (top 100%–10% vs. bottom 90%–10%). Feature importance was evaluated through recursive feature elimination (RFE) and model-based approaches, with clinically actionable thresholds also determined. Results Seven key predictors were consistently identified across sampling strategies: First-Ever Fall-Related Injury (FH1-Injury), ADL (Activities of Daily Living) score, Age, Waist Circumference, Hearing Deficit, Generalized Anxiety Disorder-7 (GAD-7) score, and the Mini-Mental State Examination (MMSE) score. Lower ADL (< 100) and lower MMSE (< 19) scores were associated with increased fall risk, reflecting functional and cognitive decline. Likewise, advanced age (≥ 79 years), higher GAD-7 (≥ 1) scores indicating anxiety symptoms, and greater waist circumference (≥ 90 cm) predicted elevated fall probability. Decision tree, AdaBoost, and gradient boosting achieved near-perfect discrimination (AUC ≈ 1.00), even when limited to the top 20% of features. Logistic regression yielded comparably high accuracy and AUC while maintaining interpretability, making it suitable for clinical deployment. Conclusions This study presents a robust and scalable ML framework that integrates multidimensional CGA data to predict falls in cognitively frail older adults. Our findings support the development of a tailored fall risk scale and inform multidimensional interventions to prevent falls in this vulnerable population.