Smart Wearable Analytics for Cycling: AI-Based Physical Exertion Prediction

We evaluated deep learning approaches for classification and regression prediction, focusing on an LSTM with Multi-Head Attention model. Data from 27 healthy participants performing cycling exercises were segmented into eight two-minute intervals. Heart rate, oxygen saturation, pedal speed (RPM), and HRV features (extracted from ECG in both frequency and time domains) served as predictive inputs. Rating of Perceived Exertion (RPE) was collected every minute and used as the predictive response, categorized into high and low exertion for classification. Physiological features and RPM from each segment were used to predict the next two-minute RPE. Feature selection via Minimum Redundancy Maximum Relevance (MRMR) and Univariate Feature Ranking (UFR) identified key predictors. The LSTM with Multi-Head Attention model achieved an MSE of 1.4 and R2 of 0.54 for regression and 82.9% accuracy with an F1 score of 86.3% for classification, demonstrating its effectiveness in exertion prediction.