Regression Analysis for Prediction of Exercise Exertion Levels Using Physiological Data

In this paper, we analyzed real-time physiological data using Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) to predict exercise exertion levels during exercise. The data was collected During a 16-minute cycling exercise for ten participants. Using wearable devices, Real-time ECG, pulse rate, oxygen saturation, pulse amplitude index (PAI), and revolutions per minute (RPM) data were collected at three intensity levels for each individual. Each subject's ratings of perceived exertion (RPE) were gathered once per minute during each exercise session. Each 16-minute cycling window was divided into eight non-overlapping windows. For each 2-minute window, heart rate, RPM, PAI, and oxygen saturation levels were averaged to form the predictive features. In addition, the heart rate variability (HRV) features were extracted by analyzing the ECG data statistically and in both time and frequency domains. The extracted features formed most of the predictive features. We used the minimum redundancy maximum relevance (mRMR) algorithm response to the collected RPE to select the best features. The leading features were then used to train and test the LSTM regression to predict the next window's exertion level.