抄録
Multiple accelerometers (chest and wrist) and physiological sensors (electrocardiogram (ECG) and electrodermal activity (EDA)) based activity classification systems have been proposed to classify those activities which are too abstruse to classify with the accelerometer-based system only. “ActiGraph” count value as well as several features of ECG and EDA were extracted from 14 persons with an open-source dataset named “PPG Field Study Dataset”. The activities were considered as different types (type A, type B, type C) based on the movement intensity of the wrist and the chest. The features were fitted to machine learning (ML) algorithms with accelerometer features only and both (count and physiological feature). In the case of type A activities (no chest and wrist movement), classification with accelerometers only and both, the accuracy was 89.3% with weighted K-NN and 100% with linear regression, respectively. In the case of classification between type B (hand soccer, cycling, driving, lunch) activities with accelerometers only and both, the accuracy was 66.1% and 89.3% with linear and quadratic support vector machine (SVM) respectively. In the case of classification between type C activities (stair and walk) with accelerometers only and both, the accuracy was 64.3% and 82.1% with linear discrimination and SVM (linear) classifier respectively. All the results for type A and type C activities classification were generated with leave one out cross-validation, and classification between type B activities was classified with 50-fold cross-validation to avoid overfitting where Bayesian optimizer tuned the hyperparameters of the ML classifiers.
