Scholarlink

Medical language understanding and understandable AI have recently made breakthrough in clinical decision support systems (CDSS). Between 2023 and 2025, large language models (LLMs), instruction-tuned models, and multimodal learning have become common to unstructured electronic health record (EHR) data analysis to make a diagnosis, treatment, and risk prediction. To secure clinician trust and regulatory compliance, however, there is need to have clear and understandable systems. The systematic review of this study is comprised of 42 peer-reviewed and preprint studies focused on the role of natural language processing (NLP) and explainable AI (XAI) in EHR-based CDSS. Such innovations as pseudo-note generation, structured-to- text pipelines, human-centered evaluation of explanations, and multimodal reasoning have been noted. Continuing issues include faithfulness in explanation, clinical validation and domain shift. We suggest standardized assessment models, increased involvement of clinicians and clinically valid standards to create credible and implementable CDSS. Objectives, In order to conclude the state of the art of NLP-based CDSS with the integration of XAI, assess the methodological advances and explainability, detect the lack of transparency and involvement of clinicians, and recommend organisational adoption strategies. Methods, In accordance with PRISMA, PubMed, IEEE Xplore, and arxiv were searched using the studies that utilized NLP on EHR text in executing CDSS tasks with explicit XAI. A total of forty-two studies were reviewed based on type of task, model category, explainability strategy and clinician evaluation and new methodological directions were identified. Results, They are transformer-based and LLM architectures (>70% of studies). The most widespread XAI technique is attention visualization, whereas the hybrid approaches to SHAP and counterfactual reasoning are gaining popularity. The research-to-practice gap was also manifested in a low proportion of 28% of studies that included clinician-in-the-loop evaluations.

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

Trust and Transparency in Healthcare AI: A Systematic Review of Explainable NLP for Clinical Decision Support (2023–2025)

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • MS Tharini,
  • Jane Rubel Angelina Jeyaraj

摘要

Medical language understanding and understandable AI have recently made breakthrough in clinical decision support systems (CDSS). Between 2023 and 2025, large language models (LLMs), instruction-tuned models, and multimodal learning have become common to unstructured electronic health record (EHR) data analysis to make a diagnosis, treatment, and risk prediction. To secure clinician trust and regulatory compliance, however, there is need to have clear and understandable systems. The systematic review of this study is comprised of 42 peer-reviewed and preprint studies focused on the role of natural language processing (NLP) and explainable AI (XAI) in EHR-based CDSS. Such innovations as pseudo-note generation, structured-to- text pipelines, human-centered evaluation of explanations, and multimodal reasoning have been noted. Continuing issues include faithfulness in explanation, clinical validation and domain shift. We suggest standardized assessment models, increased involvement of clinicians and clinically valid standards to create credible and implementable CDSS. Objectives, In order to conclude the state of the art of NLP-based CDSS with the integration of XAI, assess the methodological advances and explainability, detect the lack of transparency and involvement of clinicians, and recommend organisational adoption strategies. Methods, In accordance with PRISMA, PubMed, IEEE Xplore, and arxiv were searched using the studies that utilized NLP on EHR text in executing CDSS tasks with explicit XAI. A total of forty-two studies were reviewed based on type of task, model category, explainability strategy and clinician evaluation and new methodological directions were identified. Results, They are transformer-based and LLM architectures (>70% of studies). The most widespread XAI technique is attention visualization, whereas the hybrid approaches to SHAP and counterfactual reasoning are gaining popularity. The research-to-practice gap was also manifested in a low proportion of 28% of studies that included clinician-in-the-loop evaluations.