Explainable AI for Frailty and Fall Risk Prediction in Older Adults

Frailty is a geriatric syndrome that reflects a state of increased vulnerability to endogenous and exogenous stressors, exposing individuals to a higher risk of premature death and adverse health outcomes. This work aims at identifying new determinants of frailty and improving screening tools for less studied clinical endpoints. To this end, we analyze a novel cohort provided by the Camara Municipal de Famalicao, comprising 2,862 participants and 6,855 observations across up to four assessment moments, and spanning sociodemographics, anthropometry, functional tests, cognitive assessments, quality of life, and fall history. We combine unsupervised clustering to explore heterogeneity with supervised prediction for falls, hospitalization, and handgrip strength, using explainability approaches to connect model outputs to clinically meaningful patterns. Outcome-agnostic clustering separates functional profiles ranging from robust to vulnerable that align a posteriori with different fall burdens, while outcome-aware clustering further identifies a high-risk subgroup characterized by poorer mobility, endurance and greater reliance on mobility assistance. Supervised models achieve moderate and consistent discrimination for fall prediction (AUROC{approx} 0.66-0.68), and explainability approaches consistently emphasize key drivers including handgrip strength, self-report assessments, and other results from functional tests. Handgrip regression attains MAE{approx} 3.6 kg (R2{approx} 0.49), while a dedicated CatBoost sarcopenia classifier improves detection (AUROC = 0.798, recall = 0.792) at the cost of low precision, consistent with screening-oriented use. Overall, the results support the feasibility of explainable AI for actionable risk stratification in community assessments, while highlighting constraints related to missingness, class imbalance, and selection bias toward an active cohort.