Clustering-Based Balance Phenotyping in Older Adults from Treadmill Training Interventions

Individual heterogeneity in balance training responses among older adults remains poorly understood, as grouplevel analyses often obscure clinically meaningful variation. This study develops a multi-dimensional phenotyping framework to characterize individual responses to treadmill-based rehabilitation. Sixty older adults (mean age: 55.9 ± 5.5 years) completed an 8-week randomized controlled trial with three interventions: positive pressure with body weight support (PP-BWS), positive pressure without support (PP-noBWS), and traditional treadmill (TT). Smartphone-based sensor data were collected across six balance conditions and six time points. Five unsupervised clustering approaches were applied: (1) longitudinal trajectory clustering of temporal adaptation, (2) baseline phenotyping, (3) outcome convergence clustering, (4) condition-severity clustering of degradation under task difficulty, and (5) sensor modality clustering across accelerometer, gyroscope, and quaternion signals. Results revealed substantial heterogeneity transcending intervention groups. Temporal analysis identified three adaptation phenotypes: intervention-independent (40-50% of participants) and two intervention-sensitive patterns. Challenge profiling distinguished controlled (17.5%), intermediate (49.1%), and severe (33.3%) degradation phenotypes with slopes ranging 0.0243- 0.0430. Sensor analysis showed universal accelerometer dominance (∼ 48-50%) but underutilization of gyroscope input (∼ 7-9%), highlighting untapped therapeutic potential. One PP-BWS participant demonstrated an exceptional response, forming a unique phenotypic category (1: 58 separation), while 98.3% converged toward common outcomes. The proposed framework establishes an empirical basis for precision rehabilitation by enabling phenotype-guided intervention selection, risk-stratified progression, and targeted sensor training strategies.