Predicting Treatment Response in Body Dysmorphic Disorder with Interpretable Machine Learning

Omar Costilla-Reyes, Morgan TalbotarXiv 2025 · DOI · arXiv

Abstract: Researchers from MIT and Harvard have demonstrated that interpretable machine learning models can effectively predict treatment outcomes for Body Dysmorphic Disorder (BDD). By analyzing clinical trial data, the study reveals that a patient’s initial belief in the treatment’s effectiveness—treatment credibility—is a more potent predictor of success than baseline symptom severity. This paper explores the performance of various models and the actionable clinical heuristics derived from them.

Introduction

Body Dysmorphic Disorder (BDD) is a distressing condition affecting nearly 2-3% of the population. While Cognitive Behavioral Therapy (CBT) and SSRIs are standard treatments, many patients do not achieve remission. Clinical psychiatry faces the challenge of moving beyond "one-size-fits-all" approaches toward personalized medicine. This study addresses the need for interpretable machine learning models that provide actionable insights, identifying modifiable factors to improve recovery chances.

Methods

The study compared five machine learning approaches: Logistic Regression (LR), Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Decision Trees, and Random Forests. Data was drawn from a clinical trial of CBT for BDD, involving 17 candidate predictors such as demographics, clinical scales (BDD-YBOCS, BABS, QIDS), and treatment factors (credibility and expectancy).

Rigorous statistical techniques were employed, including multiple imputation for missing data and forward feature selection. A strict criterion ensured model parsimony: new predictors were only added if they improved the Area Under the Curve (AUC) by at least 0.05.

Model Performance and Feature Selection

The primary outcomes were Treatment Response (a $\ge$30% reduction in symptoms) and Remission (a final BDD-YBOCS score $\le$ 16). The analysis showed that complex non-linear models like Random Forests did not significantly outperform simpler linear models like SVM or Logistic Regression in this small-sample psychiatric context.

Model performance (AUC) and selected features for predicting response and remission.

As shown in the table above, the SVM model achieved the highest predictive accuracy with an AUC of 0.733 for treatment response and 0.718 for remission. Notably, treatment credibility was consistently selected as a top predictor across different model architectures.

Clinical Insights: The Power of Credibility

A central finding of the research is that treatment credibility—how much a patient believes the treatment will work—is a stronger predictor of success than initial symptom severity. This suggests that psychological readiness may be as important as the clinical state at baseline.

The Decision Tree analysis identified specific "cut-off" scores on the credibility scale (16 and 22 out of 27) that serve as practical guideposts for clinicians.

Decision tree model showing actionable heuristics for clinical decision-making.

Patients with a credibility score higher than 22 were 11 times more likely to achieve remission. Conversely, those scoring 16 or lower had significantly lower odds of responding (OR = 0.315). Furthermore, the models identified that female-identifying patients with low credibility scores were at a particularly high risk for non-response, highlighting a subgroup that may benefit from targeted motivational interventions.

Discussion and Future Work

While the study provides valuable insights, it is limited by a modest sample size and reliance on self-reported assessments. Additionally, the results establish correlation rather than causation; further research is needed to determine if actively increasing a patient's credibility score leads to improved outcomes.

Future Directions: The researchers suggest integrating these ML-driven diagnostics into digital health platforms. By identifying "at-risk" patients early, clinicians could deploy targeted "credibility-boosting" modules—such as motivational interviewing—to enhance the likelihood of a positive therapeutic outcome.

Summary updated

BibTeX
@misc{costillareyes2025predicting,
  author = {Omar Costilla-Reyes and Morgan Talbot},
  title = {{Predicting Treatment Response in Body Dysmorphic Disorder with Interpretable Machine Learning}},
  year = {2025},
  eprint = {2503.10741},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.10741},
}
Figure