Prediction Accuracy Of Sedentary Behavior Classification Model Utilizing Machine Learning

PURPOSE:To develop a machine learning-based prediction model for classifying sedentary behaviors(SBs) in a free-living using accelerometer data. METHODS:A total of 51 adults (25 males, 26 females) out of 59 participants were included, with an average age of 43.12 ± 13.42 years (Male: 42.44 ± 13.29, Female: 43.69 ± 13.71). Sedentary and physical activity data (1 min epoch) were collected by Actigraph and ActivPAL devices. Based on pre-determined sedentary activity(criterian), ActivPAL was examined the accuracy of the sedentary time. Then, Actigraph data from the lab setting was used for training eight machine learning classification models. The trained models were applied to the data set in free-living for 7 days-24 hrs and to verify the agreement between predicted and measured values of the sedentary time. The descriptive statistics and agreement indexes, including the Kappa & Pearson correlation coefficients, were calculated. All data were analyzed by Python. RESULTS:The validity of ActivPAL in measuring SB was confirmed through agreement validation (Kappa coefficient: 0.85, r: 0.86). The models trained on laboratory data showed accuracy ranging from 88.5% to 96.5% with rare differences. The predicted values in free-living were from 75.0% to 83.8% by comparing to ActivPAL, in which the SVM (Support Vector Machine) model showed the highest accuracy. CONCLUSIONS:ActivPAL was accurate enough for measuring SB, and it appears feasible to compare the model's predictions. The accuracy of the SB predictions in free-living was 83.8% utilizing a model trained on laboratory-collected data of Actigraph. Future studies should explore applying advanced models such as DNN, RNN, or LSTM to improve the accuracy of SB prediction. Supported by: National Research Foundation of Korea[NRF-2021R1A2C2013651]