Bo Zhang, Su-Wen Zhang, Tian-Qi Wu
Physical activity recovery after a population-level stressor varies substantially across individuals and cannot be adequately characterized by average step count alone. This study developed a source-aware personalized activity recovery representation for predicting short-term recovery slowdown using wearable step count data. Daily step counts from 226 participants across four source cohorts were analyzed, yielding 44,825 daily observations and 31,860 eligible participant-day prediction windows. Predictors were constructed only from observations available up to each prediction day and included baseline stability, early perturbation magnitude, recent recovery dynamics, local dynamic complexity, data-quality descriptors, and transferable source-domain descriptors. Recovery slowdown or reversal was defined using the baseline-normalized local slope of the subsequent 7-day window. This outcome was treated as a surrogate measure of short-term behavioral recovery momentum rather than as a clinically adjudicated recovery endpoint. The source-aware personalized activity recovery representation with XGBoost achieved an area under the receiver operating characteristic curve of 0.809, an area under the precision-recall curve of 0.702, balanced accuracy of 0.739, F1-score of 0.676, and Brier score of 0.162. Internal-external cross-validation yielded a mean area under the receiver operating characteristic curve of 0.771, indicating moderate cross-source transportability among cohorts exposed to the same first-lockdown event. SHapley Additive exPlanations identified recent recovery slope, local dynamic complexity, perturbation magnitude, baseline variability, and missingness ratio as key predictors. These findings establish a proof-of-concept for explainable wearable-derived early-warning research. Independent prospective validation, validation against functional or health-related outcomes, and local recalibration are required before clinical use or real-time intervention triggering.