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

Area of Science:

• Sports Science
• Biomedical Engineering
• Machine Learning

Background:

• Accurate monitoring of exercise intensity is crucial for optimizing training and preventing overexertion.
• Physiological signals and performance metrics offer potential for real-time exertion assessment.

Purpose of the Study:

• To evaluate deep learning models, specifically LSTM with Multi-Head Attention, for predicting Rating of Perceived Exertion (RPE).
• To identify key physiological and performance predictors of exertion during cycling exercise.

Main Methods:

• Collected heart rate, oxygen saturation, pedal speed (RPM), and ECG-derived HRV features from 27 participants during cycling.
• Utilized an LSTM with Multi-Head Attention model for both regression and classification (high/low exertion) of RPE.
• Employed Minimum Redundancy Maximum Relevance (MRMR) and Univariate Feature Ranking (UFR) for feature selection.

Main Results:

• The LSTM model achieved an R2 of 0.54 for regression and 82.9% accuracy with an 86.3% F1 score for classification.
• Key predictors included physiological features and RPM, highlighting their importance in exertion estimation.

Conclusions:

• Deep learning, particularly LSTM with Multi-Head Attention, demonstrates significant effectiveness in predicting perceived exertion during exercise.
• The findings support the use of physiological data and advanced machine learning for objective exercise intensity monitoring.