Scholarlink

<p>Large Language Models (LLMs) hold transformative potential for healthcare, offering capabilities, such as automated diagnosis, clinical documentation, and patient education. However, their integration into medical practice remains challenging due to critical limitations, including error propensity, bias, and privacy risks. This review provides the first systematic analysis of LLMs across diverse medical applications from electronic health record management to drug discovery while highlighting unresolved barriers to real-world adoption. We evaluate the performance of state-of-the-art models such as GPT-4, Med-PaLM2, and Bio BERT in clinical tasks, revealing disparities in accuracy, 22% hallucination rates in GPT-4 vs. 8% in Med-PaLM2 for radiology. Key challenges include biased outputs, data security vulnerabilities, and inadequate reasoning for complex medical tasks. We propose actionable strategies to mitigate these risks, such as fine-tuning with domain-specific data sets and hybrid human–AI workflows. This review bridges gaps between AI research and clinical practice, offering a roadmap for the safe, equitable, and effective deployment of LLMs in healthcare.</p>

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

Applications and limitations of large language models to integrate medical context: a comprehensive review

  • 期刊论文

  • 发表日期：

    2026年1月3日

查看全文
  • Raja Vavekanand,
  • Asif Ali Laghari,
  • Teerath Kumar

摘要

Large Language Models (LLMs) hold transformative potential for healthcare, offering capabilities, such as automated diagnosis, clinical documentation, and patient education. However, their integration into medical practice remains challenging due to critical limitations, including error propensity, bias, and privacy risks. This review provides the first systematic analysis of LLMs across diverse medical applications from electronic health record management to drug discovery while highlighting unresolved barriers to real-world adoption. We evaluate the performance of state-of-the-art models such as GPT-4, Med-PaLM2, and Bio BERT in clinical tasks, revealing disparities in accuracy, 22% hallucination rates in GPT-4 vs. 8% in Med-PaLM2 for radiology. Key challenges include biased outputs, data security vulnerabilities, and inadequate reasoning for complex medical tasks. We propose actionable strategies to mitigate these risks, such as fine-tuning with domain-specific data sets and hybrid human–AI workflows. This review bridges gaps between AI research and clinical practice, offering a roadmap for the safe, equitable, and effective deployment of LLMs in healthcare.