Robust Assessment of Free-Living Physical Behaviors and Activity Intensity Using Dual-Wearable Multitask Learning: Development and Evaluation Study From the Multicenter WEALTH Project

Background: Accurate assessment of physical behaviors (PBs) and activity intensity is essential for public health research and digital health monitoring. Wearable accelerometers combined with machine learning (ML) or deep learning (DL) enable objective behavior assessment, but most existing models are trained on laboratory data, limiting generalizability to free-living conditions. Objective: This study aimed to develop and evaluate multitask ML and DL models for PB classification across 7 categories (sitting, standing, walking, running, sports, cycling, and lying) and activity intensity categories (AIC) across 3 levels (sedentary, light, and moderate-to-vigorous physical activity) using thigh-worn (activPAL) and waist-worn (ActiGraph) wearable accelerometers. A second objective was to compare model-derived estimates of daily time spent in PB and AIC across single- and dual-sensor (activPAL + ActiGraph) configurations, and to evaluate agreement between the best-performing model and corresponding estimates obtained from the proprietary activPAL classification of real-world everyday activities (CREA) algorithm using free-living data collected over a 9-day monitoring period. Methods: Data were obtained from 590 adults in the multicenter WEALTH study (627 recruited) and included up to 9 days of concurrent activPAL and ActiGraph free-living recordings. Sparse accelerometer-labeled data were obtained using ecological momentary assessment and refined by retaining instances with ≥75% agreement with the CREA algorithm. Resulting labeled data of 583 participants were used to develop ML models for single-sensor (activPAL or ActiGraph) and combined (dual-sensor) configurations. A random forest (RF) model using engineered features and a multihead convolutional neural network (MH-CNN) were trained within a multitask learning framework to jointly predict PB (task 1) and AIC (task 2) using a subject-independent hold-out split. The test subset (n=87) was used to estimate daily time spent in PB and AIC over 9 days, which were compared with CREA-derived estimates using Pearson coefficients and intraclass correlation coefficients (ICCs). Results: The dual-sensor configuration consistently outperformed single-sensor models. For PB classification, the MH-CNN achieved the highest performance (F1-score=0.750). For AIC, the RF model performed best (F1-score=0.741). Dual-sensor free-living estimates showed epidemiologically plausible distributions across the 24-hour period, including sitting 37% (538/1440 min), lying 34% (496/1440 min), walking 9% (131/1440 min), and moderate-to-vigorous physical activity (MVPA) 2% (31/1440 min). Agreement with CREA was strongest for standing, walking, and cycling (r≥0.86; ICC ≥0.72), while lying showed modest reliability (ICC=0.48). For AIC, agreement was highest for light physical activity (LPA) and MVPA (ICC 0.72-0.75). Conclusions: Multitask models combining thigh- and waist-worn accelerometers provide consistent estimates of PB and AIC under free-living conditions. The dual-sensor approach yielded more stable and epidemiologically coherent estimates than single-sensor methods, supporting its potential for large-scale population monitoring and mobile health apps.