Explainable artificial intelligence models for accurate physical activity prediction using wearable device data.

Area of Science:

• Wearable technology and sensor data analysis
• Machine learning and artificial intelligence in health
• Biomedical signal processing

Background:

• Smartwatches offer continuous physical activity monitoring, but their algorithms lack transparency and generalizability.
• Current machine learning models for activity recognition often prioritize accuracy over interpretability.
• There is a need for explainable AI in wearable health technology to understand model decision-making.

Purpose of the Study:

• To apply explainable deep learning techniques to raw sensor data from commercial smartwatches.
• To enhance the interpretability and real-world applicability of physical activity intensity prediction models.
• To compare the performance and feature importance of different machine learning and deep learning models.

Main Methods:

• Secondary analysis of publicly available Apple Watch and Fitbit datasets.
• Training of machine learning (Decision Tree, Random Forest) and deep learning models (GRU, LSTM, CNN-LSTM).
• Classification of physical activity intensity (sedentary, light, vigorous) using sensor data.
• Assessment of model interpretability via Shapley Additive Explanations (SHAP).
• Evaluation of model performance using accuracy, precision, recall, and F1-score.

Main Results:

• The Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model achieved the highest performance (F1 scores of 82.5 for Apple Watch, 82.3 for Fitbit).
• SHAP analysis revealed distinct feature importance patterns between devices.
• Apple Watch models emphasized heart rate for vigorous activity, while Fitbit models prioritized step metrics across intensities.

Conclusions:

• Integrating deep learning with SHAP significantly improves accuracy and interpretability in physical activity intensity prediction.
• This explainable AI approach can advance wearable health monitoring systems.
• The findings support the development of real-time adaptive interventions based on personalized feedback.