Recognition of Dietary Activity via Commercial Smartwatch Orientation Sensor Analysis

Objectives: Smartwatches with movement sensors have emerged as promising tools for monitoring and improving individuals' health and lifestyle.This study focuses on applying wristwatch sensor (accelerometers and gyroscopes) data for detecting eating episodes, aiming to provide a non-intrusive, continuous monitoring of dietary behavior.Methods: Twenty healthy participants were equipped with smartwatches during four meals under semi-naturalistic settings.All meals were recorded on video and synchronized with sensors, and two reviewers manually annotated eating behavior.Raw sensor data was transformed to obtain a rectangular dataset.Rigorous machine learning methods were used to classify eating activity.Results: More than 130,000 observations were obtained and analyzed.Random Forest and Decision Tree emerged as the topperforming models.Using a comprehensive feature set, Random Forest achieved a test balanced accuracy of 0.837 and an ROC-AUC of 0.958, while Decision Tree followed closely with a test balanced accuracy of 0.826 and an ROC-AUC of 0.903.Features such as the mean and maximum (within each window) of acceleration axes significantly enhanced model performance for distinguishing between eating and non-eating episodes.K-Nearest Neighbors (KNN) and Gradient Boosting Machine (GBM) also demonstrated moderate performance, achieving test balanced accuracies of 0.808 and 0.757, respectively.Naive Bayes and Rule Fit models showed limited capability, with balanced accuracies near random performance.Conclusions: We demonstrated the efficacy of smartwatch sensors in monitoring dietary behaviors.The utilization of Random Forest and Decision Tree models highlights a promising avenue for accurately detecting eating episodes through wearable technology.This study underscores the potential of wearable devices in providing a non-intrusive and continuous approach to dietary monitoring, albeit with room for improvement in specificity metrics.