Over 40 % of adults in America experience obesity. But tackling the disease is complicated and often relies on self-reported data, which can be unreliable or inconsistent. These methods also fail to capture the dynamic, moment-to-moment nature of eating behaviors and their environmental context.
To address these gaps, researchers are exploring passive sensing technologies that offer continuous, objective monitoring. Wearable devices, such as smart cameras or physiological sensors, can enable data collection without interrupting daily life, providing a more accurate view of when, where, and how people eat.
Moving Toward Sensor-Based Behavior Profiling
The study, published in npj Digital Medicine, introduced a passive sensing framework that combines a wearable camera with a smartphone app. The camera, worn on a lanyard around the neck, captured both video and thermal imagery of participants and their immediate surroundings. The camera's Infrared sensing ensured clear imaging even in low-light conditions, and participants could pause recording at any time for privacy.
Over two years (657 days) of data collection, the researchers manually labeled more than 6,300 hours of footage to identify micromovements associated with eating, such as bites, chews, and hand-to-mouth gestures.
These behavior cues were then processed into analytical data points, including bite rate, chewing frequency, and meal duration. After preprocessing, which involved imputation of missing values, temporal smoothing, and standardization, these features were used as inputs for machine learning models to detect overeating episodes.
Machine Learning for Detection and Phenotyping
Using supervised learning techniques with cross-validation, the models achieved a mean Area Under the Receiver Operating Characteristic (AUROC) of approximately 0.86, which indicated high discriminative accuracy between overeating and typical eating behavior. Key contributing features included bite count, chew rhythm, and contextual cues such as surrounding activity and environmental setting.
The researchers also applied semi-supervised clustering using neural network encoders and dimensionality reduction techniques (notably UMAP) to identify behavioral phenotypes. This analysis revealed five distinct overeating subtypes, each characterized by unique patterns in micromovements, emotional indicators, and environmental context, all derived from sensor data alone.
By integrating the data with self-reported Ecological Momentary Assessment (EMA) information, the study provided a richer, multi-dimensional understanding of overeating triggers.
Download your PDF now!
Toward Personalized, Context-Aware Interventions
One of the most promising outcomes of this work is the potential for personalized intervention. Identifying overeating phenotypes based solely on objective sensor data could enable doctors and patients to tailor interventions based on individual behavioral profiles. For example, some subtypes may be more sensitive to environmental triggers, while others may be driven by emotional cues or habitual pacing.
Using thermal and visual data enhanced the system’s ability to detect subtle behaviors in real time, offering a foundation for future health tools that can adapt interventions based on immediate context and user behavior.
Conclusion
This study highlights the potential of wearable sensor systems to detect and analyze overeating in real-world environments objectively. By moving beyond self-reporting and incorporating continuous, high-resolution data from passive sensors, researchers have developed a framework capable of identifying overeating episodes and categorizing behavioral subtypes.
As these methods are refined and validated across diverse populations, they may significantly improve our understanding of obesity and related eating behaviors and, ultimately, how we address them.
Journal Reference
Shahabi F., et al. (2025). Unveiling overeating patterns within digital longitudinal data on eating behaviors and contexts. npj Digital Medicine. DOI: 10.1038/s41746-025-01698-9, https://www.nature.com/articles/s41746-025-01698-9