JMIR Publications

ABSTRACT

Background:

Binge- and purge–related behaviors in eating disorders are episodic and lack objective biomarkers, limiting timely detection and intervention. Wearable devices offer continuous physiological monitoring that may enable identification of physiological changes associated with symptom onset.

Objective:

This study aimed to evaluate whether wearable-derived physiological and kinematic data, analyzed using machine learning, can detect and predict binge- and purge–related states in individuals with purging-type eating disorders.

Methods:

Twenty-two adult women with purging-type eating disorders wore a commercial smartwatch continuously for two weeks, providing physiological (heart rate, heart rate variability) and kinematic data, along with time-stamped self-reports of urges and binge/purge events. Features were extracted from 30-second windows and used to train random forest classifiers for event detection and pre-event prediction. Model performance was evaluated using random splits, temporally ordered splits, and leave-one-participant-out cross-validation.

Results:

Heart rate–derived features, particularly mean heart rate and variability metrics, consistently emerged as the most informative markers. Event classification achieved an accuracy of 89.8% under temporally ordered evaluation, while prediction of imminent events yielded a mean F1-score of 0.865 and accuracy of 0.842. Substantial inter-individual variability was observed, with a mean leave-one-participant-out cross-validation F1-score of approximately 0.69.

Conclusions:

Wearable-derived physiological data can detect and predict binge- and purge–related states with clinically meaningful performance. These findings support the feasibility of wearable-based monitoring as a foundation for future just-in-time adaptive interventions in eating disorders.