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

OBJECTIVES: Wearable devices such as smartwatches are increasingly used for physical activity monitoring, but their proprietary algorithms often lack transparency and limit generalizability. Existing machine learning models prioritize accuracy, overlooking interpretability and real-world applicability. This study addresses these gaps by applying explainable deep learning techniques to raw sensor data from commercial devices. DESIGN: A cross-sectional secondary data analysis was conducted using publicly available datasets collected from Apple Watch and Fitbit devices. METHODS: Machine learning models (Decision Tree, Random Forest) and deep learning models (Gated Recurrent Unit, Long Short-Term Memory, Convolutional Neural Network-Long Short-Term Memory) were trained to classify physical activity intensity into sedentary, light, and vigorous levels. Model interpretability was assessed using Shapley Additive Explanations. Model performance was evaluated using accuracy, precision, recall, and F1-score. RESULTS: The Convolutional Neural Network-Long Short-Term Memory model demonstrated the best overall performance, achieving F1 scores of 82.5 (Apple Watch) and 82.3 (Fitbit). Shapley Additive Explanations analysis revealed device-specific patterns: Apple Watch models placed greater emphasis on heart rate-related features, particularly in vigorous physical activity, while Fitbit models prioritized step-based metrics across physical activity intensity levels. CONCLUSIONS: The integration of deep learning models with Shapley Additive Explanations enhances both predictive accuracy and interpretability in physical activity intensity level prediction. This approach may aid future health monitoring systems and support just-in-time adaptive interventions with real-time feedback.