Using unsupervised machine learning to quantify physical activity from accelerometry in a diverse and rapidly changing population

• DOI: 10.1371/journal.pdig.0000220
• Publication Date: 2023-04-05T17:39:27Z
• Authors: Christopher B. Th...; Niina Kolehmainen; Kianoush Nazarpour
• Keywords: 03 medical and health sciences; 0302 clinical medicine; FOS: Biological sciences; Computer applications to medicine. Medical informatics; R858-859.7; 10. No inequality; Quantitative Biology - Quantitative Methods; Quantitative Methods (q-bio.QM); Research Article

Accelerometers are widely used to measure physical activity behaviour, including in children. The traditional method for processing acceleration data uses cut points define intensity, relying on calibration studies that relate the magnitude of energy expenditure. However, these relationships do not generalise across diverse populations and hence they must be parametrised each subpopulation (e.g., age groups) which is costly makes over time difficult. A data-driven approach allows intensity states emerge from data, without parameters derived external populations, offers a new perspective this problem potentially improved results. We applied an unsupervised machine learning approach, namely hidden semi-Markov model, segment cluster raw accelerometer recorded (using waist-worn ActiGraph GT3X+) 279 children (9-38 months old) with range developmental abilities (measured using Paediatric Evaluation Disability Inventory-Computer Adaptive Testing measure). benchmarked analysis calculated thresholds literature had been validated same device population most closely matched ours. Time spent active as measured by correlated more strongly PEDI-CAT measures child's mobility (R2: 0.51 vs 0.39), social-cognitive capacity 0.32 0.20), responsibility 0.21 0.13), daily 0.35 0.24), 0.15 0.1) than approach. Unsupervised potential provide sensitive, appropriate, cost-effective quantifying behaviour compared current This, turn, supports research inclusive or rapidly changing populations.