Optimizing Accelerometer Placements For Detecting Upper Body Movement During Occupational Physical Activity Using Machine Learning

PURPOSE: Occupational physical activity (OPA) may have a distinct, potentially adverse, relationship with health. Thigh-worn activPAL accelerometers, common in free-living physical activity measurement, do not adequately detect upper body movements typical to OPA. This study aimed to identify the most accurate and feasible accelerometer placement(s) and classification models for detecting OPA-related upper body movement. METHODS: 19 participants (age = 35.3 ± 13.4 years, 47% female, 74% white) completed 11 laboratory-simulated OPA tasks with and without upper body movement for 3 minutes each: sitting, seated typing, seated assembly, seated painting, standing, standing holding a box, standing assembly, box lifting, walking, walking carrying, and mopping. Participants wore one activPAL on their dominant thigh (T), upper back (B), upper arm (A), and wrist (W), respectively. Raw triaxial accelerometer signals were low-pass filtered, summarized in 1-second epochs, and split into 25% test/75% train datasets. Second-by-second classification accuracy of upper body movement or not was examined using three machine learning models (logistic, random forest, and gradient boosted) and nine sensor placement combinations (T only, A only, B only, W only, T + B, T + W, T + A, T + A + W, and T + B + A + W). RESULTS: Table 1 summarizes the test data classification accuracy across sensor placements model types. Random forest models consistently had the highest accuracy (89.38-99.97%). T only provided the lowest accuracy (54.07-89.38%). W only and models which included at least one A, B, or W sensor along with T (T + B, T + A, T + W, T + A + W, and T + B + A + W) all resulted in ≥99% accuracy. CONCLUSIONS: Upper body movement during laboratory-simulated OPA can be accurately classified using a single W activPAL or by combining W, A, or B with T which also measures lower body ambulation/posture. After field validation, such methods may more accurately measure occupational PA exposure for health.The project described was supported by the National Institute of General Medical Sciences, 2 U54GM104942-07