Exercise Exertion Levels Prediction

This pilot study used Long Short-Term Memory (LSTM) and gated recurrent unit (GRU) Recurrent Neural Network (RNN) models to predict exercise exertion levels based on physiological data acquired from wearable devices. The data, including revolutions per minute (RPM), level of oxygen saturation, real-time ECG, and pulse rate were collected during a 16-minute cycling exercise across three intensity levels. Ratings of perceived exertion (RPE) from study subjects were recorded at the end of each consecutive minute during every session. The 16-minute sessions were segmented into eight 2-minute windows, and each window was categorized as either "high exertion" or "low exertion" based on self-reported RPEs. Predictive features were calculated by averaging oxygen saturation levels, heart rate, and RPMs for each 2-minute window. Additionally, heart rate variability (HRV) features were extracted from collected ECG data in both temporal and frequency domains. The selection of predictive features was performed using the minimum redundancy maximum relevance (mRMR) algorithm. The top features selected were employed to train and test the LSTM and the GRU classifiers for predicting the exertion level of the subsequent window. The resulting classifiers exhibited testing accuracies of 78.6% and F-1 scores of 80%. This pilot study underscores the potential of employing a deep learning model for real-time prediction of perceived exercise exertion.