Advancing fall risk prediction in older adults with cognitive frailty: A machine learning approach using 2-year clinical data

Abstract

Falls are a critical concern in older adults with cognitive frailty (CF). However, previous studies have not fully examined whether machine learning models can predict falls in older individuals with CF. The 2-year longitudinal data set from the Korean Frailty and Aging Cohort Study and machine learning approach were utilized to predict fall risk. We analyzed multidimensional health data, including demographics, clinical conditions, as well as the physical and psychological health factors of 443 older adults with CF identified out of 2,404 older adults. For fall risk prediction, we developed a machine learning framework incorporating logistic regression, bootstrapping, and recursive feature elimination. Statistical analysis revealed significant differences between the non-faller and faller groups for nine clinical conditions as well as physical and psychological variables. Using nine significant variables, our machine-learning-based model demonstrated good predictive performance with an area under the curve (AUC) exceeding 80%. Furthermore, our machine learning framework identified four optimal variables: the number of Fried physical frailty (PF) phenotypes, PF-Mobility scores, scores from the Korean version of the Short Geriatric Depression Scale, and scores from SARC-F (consisting of five components: strength, assistance with walking, rising from a chair, climbing stairs, and experiencing falls). It demonstrated excellent predictive performance, with an AUC, sensitivity, specificity, and accuracy exceeding 95%. These variables reflect the critical association between physical and psychological health and fall risk. These findings underscore the importance of integrating multidimensional health data with machine learning methodologies to accurately predict fall risk in older adults with CF, design targeted interventions, and enable healthcare professionals to implement strategies to reduce and prevent such falls.

Fig 1. STROBE diagram of eligibility determination and participant group classification.

Fig 2. Machine learning framework.

AUC, area under the curve; SMOTE, synthetic minority oversampling technique.