Modeling Perceived Exertion with Deep Neural Networks and Wearable Sensors

This study explores deep learning methods for predicting physical exertion using physiological data acquired from wearable sensors. We evaluated two advanced architectures: a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model and an LSTM model enhanced with a Multi-Head Attention mechanism. Data were collected from 27 healthy participants during structured 16-minute cycling sessions, where ECG, heart rate, oxygen saturation, and pedal speed (RPM) were continuously monitored. Each session was divided into eight two-minute segments, and heart rate variability (HRV) features were extracted for model input. Perceived exertion levels, assessed using the Rating of Perceived Exertion (RPE), were recorded at one-minute intervals. For the classification task, physiological features from each two-minute window were used to predict the exertion category (high vs. low) for the subsequent interval. Feature selection techniques, including Minimum Redundancy Maximum Relevance (MRMR) and Univariate Feature Ranking (UFR), were employed to optimize model inputs. Both models were trained and evaluated independently using block-wise cross-validation to support temporal generalization. The CNN-LSTM model with UFR-based features achieved the best classification results, with an accuracy of 85.7%, an F1-score of 88.9%, and an AUC of 89.5%. In the regression task, the objective was to predict continuous RPE scores from the physiological data. The LSTM with Multi-Head Attention model, leveraging MRMR-selected features, reached a mean squared error (MSE) of 1.4 and an R² value of 0.54. These findings highlight the potential of deep learning for real-time exertion monitoring in personalized fitness, sports performance analysis, and rehabilitation settings.