Dynamic Prediction of Physical Exertion: Leveraging AI Models and Wearable Sensor Data During Cycling Exercise.

Area of Science:

• Physiological computing
• Wearable sensor technology
• Machine learning in sports science

Background:

• Predicting physical exertion is crucial for optimizing training and preventing overexertion.
• Wearable devices offer a non-invasive method for collecting continuous physiological data.
• Existing methods for exertion prediction often lack accuracy and real-time adaptability.

Purpose of the Study:

• To investigate the efficacy of machine learning (ML) and deep learning (DL) models for predicting physical exertion.
• To evaluate the performance of various feature selection algorithms in identifying key physiological indicators.
• To develop predictive models using data from wearable sensors during controlled exercise.

Main Methods:

• Collected physiological data (ECG, heart rate, oxygen saturation, RPM) from 27 participants during cycling.
• Calculated Heart Rate Variability (HRV) as a predictive feature.
• Employed Minimum Redundancy Maximum Relevance (MRMR) and Univariate Feature Ranking for feature selection.
• Developed 34 traditional ML models and Long Short-Term Memory (LSTM) networks for classification and regression tasks.

Main Results:

• The LSTM regression model achieved a Mean Squared Error (MSE) of 0.8493 and R-squared of 0.7757.
• Classification models reached a maximum testing accuracy of 89.2% and an F1 score of 91.7%.
• Feature selection algorithms effectively identified relevant physiological signals for model training.

Conclusions:

• Combining feature selection with ML and DL significantly enhances physical exertion prediction accuracy.
• Wearable sensor data, when analyzed with advanced algorithms, provides reliable insights into exertion levels.
• This approach holds promise for personalized fitness monitoring and performance optimization.