Generalizable Physical Therapy Exercise Classification from Wearable IMU Data Using LOSO Evaluation

Abstract

In this work, we propose a system to recognize human activities using the accelerometer and gyroscope of a smartphone, which has the potential to be used as a tool to monitor health conditions in real time. We use the UCI HAR dataset to assess the performance of different classifiers, such as Decision Tree, Naive Bayes, SVM, Random Forest, ANN, and XGBoost, as well as the effect of window sizes, as well as the generalization ability of the proposed system using the Leave-One-Subject-Out method. The ensemble classifiers, such as the XGBoost model, have the best performance among the classifiers used, with a maximum F1 score of 98.1%, as shown by the feature importance plot, which indicates the effectiveness of the combined time- and frequency-domain features used in the proposed