Comparison of classification accuracy for machine learning models in types of squat motion: focusing on normal/abnormal squat

The purpose of this study is to evaluate the classification accuracy applying for machine learning model for normal/abnormal squat. One hundred participants (age: 24.2±2.4 years, height: 177.4±5.1 cm, body mass: 73.3±7.1 kg) were included in this study and collected 12,000 for normal/abnormal squat data. During squat, angle of trunk, hip, knee ankle joint, angle difference of left and right, position difference of shoulder, elbow, wrist joint. To automatically classify squat movements, principal component analysis (PCA), k-nearest neighbors (kNN), logistic regression (LR), support vector machine (SVM), decision tree (DT), and random forest (RF) models were used, and the variables were ranked based on feature importance. The results of classifying the normal squat against the abnormal squat were as follows: kNN 0.975, LR 0.841, SVM 0.865, RF 0.977, and DT 0.960. For abnormal squat 1, the accuracy of kNN was 0.926, LR 0.878, SVM 0.892, RF 0.997, and DT 0.997. For abnormal squat 2, the accuracy of kNN was 0.958, LR 0.887, SVM 0.898, RF 0.989, and DT 0.978. For abnormal squat 3, the accuracy of kNN was 0.915, LR 0.827, SVM 0.895, RF 0.980, and DT 0.950. The results of feature importance that normal squat were ranked in the order of Spd, Epd, and LKa. Abnormal squat 1 and 2 was rankled in the order of Tay and Epd. Abnormal squat 3 was rankled in the order of LKa, Tay, RAa, and LAa. Therefore, it is expected that using the RF machine learning model with a focus on the vertical position difference between the shoulder and elbow joints would sufficiently detect injury and accident during squat.