Reproduce “MDD-Thinker: Towards Large Reasoning Models for Major Depressive Disorder Diagnosis” on OpenResearch

About this paper

Yuyang Sha, Hongxin Pan, Gang Luo, Caijuan Shi, Jing Wang, Kefeng Li

MDD-Thinker, an LLM-based framework from Macao Polytechnic University and collaborators, was developed for Major Depressive Disorder diagnosis, leveraging a large-scale real-world clinical dataset from the UK Biobank. It integrates Supervised Fine-Tuning and Reinforcement Learning to achieve high diagnostic accuracy (AUROC of 0.8803) while generating explicit, multi-step reasoning paths for enhanced interpretability.

Key insights

• The combination of SFT and RL is complementary and highly effective, with SFT establishing core domain knowledge and RL significantly refining the inference process for both accuracy and interpretability.
• Advanced Chain-of-Thought prompting is critical for transforming LLMs into transparent decision-support tools by enabling the generation of explicit, logically coherent reasoning paths.
• Training on large-scale real-world clinical data, like the UK Biobank, is crucial for developing robust, reliable, and domain-specific LLMs for medical diagnosis, mitigating biases often present in synthetic datasets.

Summary by alphaXiv.