Impact of Sensor-Axis Combinations on Machine Learning Accuracy for Human Activity Recognition Using Accelerometer Data in Clinical Settings

Background: Identifying patient activity during symptom onset is crucial in conditions such as chronic obstructive pulmonary disease (COPD) or arrhythmia. Accelerometer data can aid in this process; however, minimal data requirements are preferable. This study aimed to determine the minimum number of axes required to accurately recognize activities related to COPD or arrhythmia. Methods: Thirty healthy participants wore a 9-axis accelerometer in five positions and performed nine activities. Machine learning-based activity recognition was conducted using 9-, 6-, and 3-axis data from the nondominant wrist or chest, as these two positions demonstrated high recognition accuracy in our previous study. Results: For the nondominant wrist, the accuracy of recognizing activities such as lying in the supine position, standing, eating, and running was comparable for the 3-axis acceleration and 9-axis data. For the chest, the accuracy of recognizing activities such as lying in the supine position, standing, sitting, using the restroom, and ascending or descending stairs was comparable for 6-axis (combined acceleration and magnetometer) and 9-axis data. Conclusions: The findings suggest that 3-axis acceleration data for the nondominant wrist and 6-axis combined acceleration and magnetometer data for the chest lead to high activity recognition accuracy.