A Machine Learning Approach for Predicting Physical Activity Intensity from Wearable Sensor Data

The paper aims to establish a robust machine learning approach for accurately categorizing the level of physical activity based on data collected from wearable sensors. Data from more than 10,000 participants was gathered through the utilization of sensors embedded in smartphones and smartwatches. These sensors include accelerometers, gyroscopes, and heart rate monitors. The acquisition of critical attributes, such as heart rate metrics, spectral entropy, and variance, involves meticulous feature engineering and thorough preprocessing. These measures effectively capture the temporal and spectral aspects of physical activity. An ensemble model is constructed using Gradient Boosting Machines (GBM) to achieve precise predictions of activity intensity categories. The GBM model consistently maintains accuracy levels above 99% for all types of physical activity, ranging from moderate to vigorous, sedentary, and mild. In order to ensure the model's robustness and prevent overfitting, hyperparameter optimization is conducted through k-fold cross-validation. The findings demonstrate the impressive precision with which machine learning algorithms classify activities. The implications can have far-reaching effects on the promotion of physical activity and the improvement of health outcomes.