Obesity Abstracts

Background: Participation in physical activity (PA) and avoidance of sedentary behaviours (SB) are important factors in the prevention of obesity. Self-report PA measures are subject to misreporting and therefore objective, accurate measurements are required. Machine learning (ML) applied to physiological and accelerometery data may offer a means to improve the classification of PA.

Methods: Subjects (n=59) were recruited to participate in a protocol consisting of resting, ambulatory, running, cycling and household tasks, whilst wearing an accelerometer (Actigraph GT3X) and a heart rate sensor (Polar H7). After incomplete data were removed, 2049 min (n=55) of steady state accelerometer, heart rate and participant data were to available to develop ML models. Algorithms included: Naïve Bayes (NB), k-Nearest Neighbour (k-NN), Support Vector Machines (SVM), Random Forest (RF) and Artificial Neural Networks (ANN). The developed models were trained to predict i) activity type and ii) activity intensity category, defined by metabolic equivalents (METs) (light: <3, moderate 3–6 & vigorous >6). ML models were validated with a leave-one-out cross validation approach.

Results: To predict activity type, ANN, K-NN, SVM and RF exceeded 95% F1-score and k-NN was 100% accurate for the classification of SB. To classify MET category, all models surpassed the accuracy of the Sensewear Armband (70.8%), with k-NN having the highest F1-score (94.2%).

Conclusion: The developed algorithms demonstrate a high degree of accuracy for the classification of type and intensity of PA. ML algorithms outperform a widely used and validated research-grade device.

Keywords: Machine learning, physical activity, sedentary behaviour, accelerometer