加齢に伴う身体機能障害のスケーラブルなデジタル表現型を可能にする自発的スピーチ

The rising global burden of pathological aging engenders an urgent need for accessible tools enabling early detection of physical decline, which significantly impacts quality of life and healthcare systems. We hypothesized that speech analysis could capture phenotype-specific signatures of physical deterioration through shared neuromuscular pathways, offering a novel approach to physical assessment. In this study, we employed machine learning to analyze multimodal speech features (acoustic, linguistic, temporal) derived from two 1-minute spontaneous emotional speech recordings obtained from 271 community-dwelling older adults (mean age: 77.3 ± 5.8 years). Our models classified physical functional deficits across ten critical domains: lower-limb strength, power, endurance, handgrip strength, flexibility, postural balance, gait speed, mobility, appendicular lean mass, and fatigue. Our ensemble approach achieved remarkable classification accuracy for each domain (mean AUC = 0.91 ± 0.04), with multimodal emotional task stacking enhancing detection for 80% of physical measures. Explainable AI (SHAP) analysis revealed distinct speech signatures for each deficit type, potentially reflecting specific pathophysiological mechanisms rather than demographic confounders. We identified three primary speech alteration clusters: lexico-syntactic simplification (decreased syntactic complexity), neuromotor-temporal slowing (diminished speech rate, increased pauses), and articulatory-spectral decline (spectral instability). This study supports the hypothesis that spontaneous speech serves as a comprehensive digital biomarker of multidimensional physical function in aging. Our approach pioneers speech analysis as a physical aging clock. This technology offers clinical-grade precision through accessible smartphone recordings, enabling domain-specific physiological mapping via interpretable biomarkers and scalable screening for precision geriatrics and underserved populations.