Eating detection from smartwatch sensors under subject-independent validation
Abstract
Abstract Background Accurate dietary assessment is essential for preventing non-communicable diseases, yet self-reported methods are unreliable. Wearable devices offer a scalable alternative, but most studies rely on research-grade sensors and within-subject validation, leading to overoptimistic results. Methods We evaluated the ability of a commercial smartwatch (Garmin Fenix 5) to detect eating episodes using accelerometer and orientation data. Twenty healthy adults were recorded during four standardized meals in semi-naturalistic settings. Data were segmented into time windows and transformed into statistical features. Multiple machine learning and deep learning models were trained and evaluated using Leave-One-Subject-Out cross-validation to ensure subject-independent performance. Results A total of 26,304 observations from 19 participants were analyzed. The best-performing model (tuned XGBoost) achieved a balanced accuracy of 0.642 (95% CI 0.612–0.671) and AUC of 0.712 (0.671–0.749). Simpler models, including logistic regression, showed comparable performance. A Transformer model achieved the highest sensitivity (0.691 [0.637–0.735]) but with reduced specificity (0.504 [0.456–0.549]). Substantial inter-subject variability was observed, with sensitivity ranging widely across individuals. Conclusions Eating detection using commercial smartwatch sensors remains challenging under realistic subject-independent validation. Performance is lower than previously reported studies using less rigorous designs, highlighting the importance of proper validation. High inter-individual variability limits generalizability, suggesting that personalization and larger datasets are necessary for real-world deployment. Key messages • Commercial smartwatches can detect eating episodes, but performance drops under subject-independent validation, highlighting real-world limitations. • High inter-individual variability limits generalizability; personalization and larger datasets are essential for reliable dietary monitoring with wearables.