Detecting Binge Eating Risk With Naturalistic Data and Machine Learning: A Comparative Observational Study.

Area of Science:

• Digital health
• Behavioral science
• Machine learning

Background:

• Negative affect is a key trigger for binge eating (BE).
• Just-in-time adaptive interventions (JITAIs) aim to reduce maladaptive behaviors by prompting therapy skills.
• Ecological momentary assessment (EMA) via smartphone surveys is the current standard for identifying negative affect but has limitations in temporal granularity, participant insight, and adherence.

Purpose of the Study:

• To compare the accuracy of machine learning models in predicting BE using EMA data versus wearable sensor data.
• To explore the potential of passively collected physiological data for detecting risk of negative affect-related BE.

Main Methods:

• Thirty adults with recurrent BE participated in a 4-week study.
• Participants wore smartwatches to collect heart rate and electrodermal activity data.
• Affect and BE were self-reported using EMA.
• Machine learning models (support vector machines, random forest, neural network) were trained and evaluated.

Main Results:

• Sensor-only models achieved a macro-averaged accuracy of 0.64 (AUROC = 0.78), comparable to EMA-only models (macro-averaged accuracy = 0.64, AUROC = 0.69).
• Sensor-only models showed higher sensitivity (0.92) but lower specificity (0.35) compared to EMA-only models.
• Combined EMA and sensor data models did not significantly improve prediction accuracy over sensor-only or EMA-only models.

Conclusions:

• Psychophysiological data from wearable sensors can predict BE with accuracy comparable to traditional EMA.
• This research paves the way for developing low-burden JITAIs that leverage wearable sensor data to target negative affect-driven BE.
• Passive sensing offers a promising alternative to overcome the limitations of self-report-based EMA for real-time behavioral interventions.