Predicting the functional response to physical intervention from speech: a validated Machine Learning model for personalized aging trajectories

Predicting individual responses to physical interventions is a critical unmet need in precision geriatric medicine. Speech production integrates respiratory, neuromotor, and cognitive systems that parallel age-related physiological decline, making it a candidate digital biomarker of functional reserve. In a longitudinal cohort of 464 community-dwelling older adults (mean age 79.6±8.7 years; 66.6% female) enrolled in a 12-week adapted physical activity program, we developed machine learning models to predict individual response trajectories across ten physical domains. Baseline speech recordings from one-minute autobiographical recall tasks yielded 128 acoustic, temporal, and linguistic features. Response trajectories were stratified into five ordinal categories, from deterioration to improvement. Models were developed using Random Forest, XGBoost, SVM, and logistic regression, with stacking ensembles combining emotional modalities. An independent chronological external cohort (n=52) was reserved for validation. Speech-based models achieved robust predictive accuracy, with Matthew's Correlation Coefficients ranging from 0.774 (95% CI 0.71-0.82) to 0.985 (95% CI 0.96-0.99) on internal test sets. ROC AUC values exceeded 0.96 for all domains, and precision-recall AUC ranged from 0.870 to 0.988. External validation confirmed generalizability with stable performance and practical equivalence between cohorts (TOST p<0.05). SHAP analysis revealed temporal speech features dominated predictive importance (48-66%), reflecting shared neural substrates of motor control. Acoustic features captured respiratory-laryngeal integrity for strength domains and linguistic features indexed cognitive reserve under fatigue. Machine learning models applied to spontaneous speech predict individual intervention response trajectories. This accessible, low-cost approach enables decentralized trajectory anticipation and personalized intervention allocation, representing a scalable digital biomarker for precision geriatric medicine. Trial registration: ClinicalTrials.gov NCT02343286.