When PCOS Meets Eating Disorders: An Explainable AI Approach to Detecting the Hidden Triple Burden

TL;DR

This paper presents explainable AI models trained on social media data to detect the co-occurrence of PCOS, body image issues, and eating disorders, emphasizing transparency and interpretability.

Contribution

It introduces small open-source language models fine-tuned for explainability to identify the triple burden in social media posts, addressing transparency gaps in NLP approaches.

Findings

Best model achieved 75.3% accuracy on held-out posts.

Models effectively detected comorbid conditions with strong explainability.

Performance decreased with increased diagnostic complexity.

Abstract

Women with polycystic ovary syndrome (PCOS) face substantially elevated risks of body image distress, disordered eating, and metabolic challenges, yet existing natural language processing approaches for detecting these conditions lack transparency and cannot identify co-occurring presentations. We developed small, open-source language models to automatically detect this triple burden in social media posts with grounded explainability. We collected 1,000 PCOS-related posts from six subreddits, with two trained annotators labeling posts using guidelines operationalizing Lee et al. (2017) clinical framework. Three models (Gemma-2-2B, Qwen3-1.7B, DeepSeek-R1-Distill-Qwen-1.5B) were fine-tuned using Low-Rank Adaptation to generate structured explanations with textual evidence. The best model achieved 75.3 percent exact match accuracy on 150 held-out posts, with robust comorbidity detection and strong explainability. Performance declined with diagnostic complexity, indicating their best use is for screening rather than autonomous diagnosis.