Predicting Real-Time Exercise Exertion with Deep Learning: Insights from Wearable Device Data

This research utilized physiological signals gathered from wearable devices to forecast levels of physical exertion during exercise using deep learning methods for classification and regression. The study involved ten healthy participants who engaged in cycling exercises lasting 16 min. Data such as ECG, heart rate, oxygen levels, and pedal speed (RPM) were monitored across varying intensities throughout each session. Additionally, participants self-reported their exertion levels at one-minute intervals. These sessions were segmented into eight parts of two minutes each, during which averages of heart rate, RPM, oxygen saturation, and reported exertion levels were calculated for model training. Furthermore, we extracted heart rate variability (HRV) as additional predictive indicators. Employing several feature selection techniques, we identified the most impactful predictors for model training. Our models demonstrated significant accuracy and precision, with training accuracies and F1 scores peaking at 98.2% and 98%, and testing performances achieving a top accuracy and F1 score of 80%.