Predicting Ratings of Perceived Exertion during Cycling Exercise Using Physiological Data from Wearable Sensors : LSTM Application

In this research, physiological data collected from wearable devices were leveraged to create deep-learning models for predicting levels of exercise exertion. A group of twenty-one healthy participants took part in 16 -minute cycling exercises, during which real-time data on ECG, pulse rate, oxygen saturation, and revolutions per minute (RPM) were recorded at three different intensity levels. Participants reported their ratings of perceived exertion (RPE) once every minute. The 16-minute cycling exercise was conducted across three intensity levels: two minutes at low intensity without resistance, four minutes at medium resistance, and ten minutes at high intensity with maximum resistance. Each intensity level was divided into consecutive 2 -minute segments, resulting in eight segments per session. During each segment, average values for RPE, heart rate, RPM, and oxygen saturation were calculated and used as predictive features. Additionally, heart rate variability (HRV) metrics were derived from the ECG data. The MRMR (Minimum Redundancy Maximum Relevance) feature selection method was utilized to determine the most relevant predictors for both classification and regression tasks. These selected features were then used to train and assess LSTM (Long Short-Term Memory) models. The classification model attained an accuracy of 90% and an F1-score of 0.89 on the test data, while the regression model achieved a mean squared error (MSE) of 1.1 on the test data.