Scholarlink

<p>The widespread use of machine learning (ML) in neurocognitive research has been limited by the black-box nature of many predictive models, which decreases trust, interpretability, and clinical application. Explainable artificial intelligence (XAI) methods are being developed to address this issue by providing both model-agnostic and model-specific explanations of algorithmic decisions. While previous work has mainly concentrated on imaging and text data, applying XAI to tabular and numerical clinical datasets remains relatively underexplored. This systematic review compiles evidence from 27 peer-reviewed studies (2015–2025) that used XAI with tabular and numerical data for neurocognitive disorders, including Alzheimer’s disease (AD), Parkinson’s disease, attention-deficit/hyperactivity disorder (ADHD), autism spectrum disorder (ASD), and stroke, as well as related psychiatric conditions like depression and anxiety that have overlapping cognitive or neural features. The analysis reveals that tree-based ensembles such as Random Forests and XGBoost are predominant in predictive modeling, often paired with SHapley Additive exPlanations (SHAP) to measure local and global feature importance. Methods like Local Interpretable Model-agnostic Explanations (LIME), rule-based models, and surrogate learning frameworks were also used, but less frequently. Across tasks including diagnosis, prognosis, classification, and comorbidity analysis, XAI improved transparency, provided feature-level insights, and offered clinically relevant interpretations of biomarkers. Nonetheless, there are still methodological and translational challenges, such as limited dataset sizes, imbalanced samples, a lack of standardized evaluation for explanation quality, and minimal external validation. Future directions involve developing fidelity metrics for explanations, integrating multimodal and longitudinal datasets, exploring hybrid pipelines that balance accuracy and interpretability, and embedding XAI into clinician-facing decision-support systems. By focusing on tabular clinical data across multiple disorders, this review complements earlier imaging-focused surveys. It emphasizes the crucial role of XAI in creating trustworthy, generalizable, and clinically aligned machine learning solutions.</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

Application of Explainable Artificial Intelligence for Predicting and Diagnosing Neurocognitive Disorders Using Tabular Data: A Systematic Review

  • 期刊论文

  • 发表日期：

    2025年12月8日

查看全文
  • Hamideh Gholipour Lazarjani,
  • Soheil Zarei,
  • Foad Aboutorabi,
  • Reza Shalbaf,
  • Ahmad Shalbaf

摘要

The widespread use of machine learning (ML) in neurocognitive research has been limited by the black-box nature of many predictive models, which decreases trust, interpretability, and clinical application. Explainable artificial intelligence (XAI) methods are being developed to address this issue by providing both model-agnostic and model-specific explanations of algorithmic decisions. While previous work has mainly concentrated on imaging and text data, applying XAI to tabular and numerical clinical datasets remains relatively underexplored. This systematic review compiles evidence from 27 peer-reviewed studies (2015–2025) that used XAI with tabular and numerical data for neurocognitive disorders, including Alzheimer’s disease (AD), Parkinson’s disease, attention-deficit/hyperactivity disorder (ADHD), autism spectrum disorder (ASD), and stroke, as well as related psychiatric conditions like depression and anxiety that have overlapping cognitive or neural features. The analysis reveals that tree-based ensembles such as Random Forests and XGBoost are predominant in predictive modeling, often paired with SHapley Additive exPlanations (SHAP) to measure local and global feature importance. Methods like Local Interpretable Model-agnostic Explanations (LIME), rule-based models, and surrogate learning frameworks were also used, but less frequently. Across tasks including diagnosis, prognosis, classification, and comorbidity analysis, XAI improved transparency, provided feature-level insights, and offered clinically relevant interpretations of biomarkers. Nonetheless, there are still methodological and translational challenges, such as limited dataset sizes, imbalanced samples, a lack of standardized evaluation for explanation quality, and minimal external validation. Future directions involve developing fidelity metrics for explanations, integrating multimodal and longitudinal datasets, exploring hybrid pipelines that balance accuracy and interpretability, and embedding XAI into clinician-facing decision-support systems. By focusing on tabular clinical data across multiple disorders, this review complements earlier imaging-focused surveys. It emphasizes the crucial role of XAI in creating trustworthy, generalizable, and clinically aligned machine learning solutions.