Leveraging artificial intelligence to personalize treatment for eating disorders: A proof-of-concept study.

Eating disorders are complex conditions with high relapse and mortality rates, and fewer than half of adults achieve clinically significant improvement with current evidence-based treatments. Personalized approaches using idiographic, data-driven monitoring have emerged to reveal individual symptom profiles, yet existing methods struggle with high-dimensional data and lack guidance for adapting care in real time. Artificial intelligence (AI) offers a complementary strategy by identifying explainable patterns in complex datasets, supporting more precise and adaptive interventions. We collected ecological momentary assessment (EMA) data from 35 participants with eating disorder diagnoses. Baseline consisted of five EMA surveys per day for 15 days, followed by two surveys per day during 12-13 weeks of treatment, yielding thousands of time-points across 14 weeks. EMA items assessed behaviors, cognitions, affect, and co-occurring symptoms on a 0-100 scale. Responses were fuzzy-encoded into graded categories to address skew and missingness. Dimensionality reduction was performed with a self-supervised Auto-Encoder, and fuzzy c-means clustering of latent representations was projected back onto EMA features to generate interpretable symptom phenotypes anchored in eating anxiety. Treatment response was quantified using Mahalanobis distance between successive embeddings, a covariance-sensitive metric highlighting atypical psychological shifts over time. The Auto-Encoder achieved RMSE = 0.25 and R