Scholarlink

The rise of virtual healthcare systems, evolving from rule-based models to Artificial Intelligence (AI), has significantly improved access to medical guidance. The emergence of Large Language Models (LLMs) has shown promising capabilities in medical diagnostics. However, the effectiveness of LLMs heavily depends on the quality and comprehensiveness of input data. This dependency presents challenges in achieving accurate diagnoses. To address this, we propose an adaptive Diagnostic Optimization Learning Framework (DOLF) that integrates LLMs with Reinforcement Learning (RL). The RL component employs a modified Q-learning algorithm to dynamically select the most informative questions based on patient responses. This approach reduces the number of required interactions while effectively collecting relevant symptom information. The collected symptom information is structured into a detailed prompt for the LLM. This ensures a comprehensive and context-rich input that leverages the LLMs extensive medical knowledge. Experimental evaluations on a curated dataset show that DOLF outperforms standalone LLM-based diagnosis. It achieves an accuracy of 94% with an average dialogue length of 4.26 turns. These results make remote diagnosis more precise and pave the way for scalable, AI-driven medical assistants to enhance global healthcare accessibility and efficiency.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Adaptive Reinforcement Learning Framework for Enhancing LLM-Based Medical Diagnostics

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mostafijur Rahaman,
  • Sounak Banerjee,
  • Sarmistha Neogy,
  • Sarbani Roy

摘要

The rise of virtual healthcare systems, evolving from rule-based models to Artificial Intelligence (AI), has significantly improved access to medical guidance. The emergence of Large Language Models (LLMs) has shown promising capabilities in medical diagnostics. However, the effectiveness of LLMs heavily depends on the quality and comprehensiveness of input data. This dependency presents challenges in achieving accurate diagnoses. To address this, we propose an adaptive Diagnostic Optimization Learning Framework (DOLF) that integrates LLMs with Reinforcement Learning (RL). The RL component employs a modified Q-learning algorithm to dynamically select the most informative questions based on patient responses. This approach reduces the number of required interactions while effectively collecting relevant symptom information. The collected symptom information is structured into a detailed prompt for the LLM. This ensures a comprehensive and context-rich input that leverages the LLMs extensive medical knowledge. Experimental evaluations on a curated dataset show that DOLF outperforms standalone LLM-based diagnosis. It achieves an accuracy of 94% with an average dialogue length of 4.26 turns. These results make remote diagnosis more precise and pave the way for scalable, AI-driven medical assistants to enhance global healthcare accessibility and efficiency.