Scholarlink

<p>With over ten million individuals affected globally, Parkinson’s disease (PD) remains noto-riously difficult to diagnose early because clinicians still depend largely on subjective motor examinations. This process often delays treatment until substantial dopaminergic loss has already occurred. In this paper we describe a multiagent Retrieval-Augmented Generation (RAG) platform that automates PD stage prediction directly from T1-weighted MRI. The pipeline runs on edge hardware under a hybrid cloud arrangement: an EfficientNet-B0 clas-sifier performs on-device inference while three cooperating agents (a Supervisor, an AI/ML module, and a RAG module) coordinate through shared memory to generate citation-backed clinical reports. Training and evaluation used the PPMI cohort (39,534 slices drawn from 1,064 subjects). On this data the classifier reaches 99.62% accuracy at the scan level and 91.4% under subject-level five-fold cross-validation, spanning six categories (healthy con-trol plus Stages 1–5). Coupling the classifier with the RAG agent cuts large-language-model hallucination from 18.5% down to 2.0% and lifts citation accuracy from 12% to 96%. A component-wise ablation confirms that each agent makes a distinct, measurable contribu-tion. Grad-CAM heatmaps, validated against atlas-defined neuroanatomical regions, show statistically significant overlap with structures known to be affected in PD. Finally, a 100-case de-identified workflow simulation yielded a 40% reduction in documentation time together with a clinician–system agreement of Cohen’s <InlineEquation ID="IEq10"><EquationSource Format="TEX">\(\bf\kappa\)</EquationSource></InlineEquation> = 0.92.</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

Multi-Agent Retrieval-Augmented Generation System for Automated Parkinson’s Disease Stage Prediction: An Edge-Computing Approach with Explainable AI

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月21日

查看全文
  • V. R. Tejas,
  • Sagar Ganiga,
  • N. R. Venkatesh Raju,
  • Safwan Sayeed,
  • Hema M S,
  • Nagashettappa Biradar

摘要

With over ten million individuals affected globally, Parkinson’s disease (PD) remains noto-riously difficult to diagnose early because clinicians still depend largely on subjective motor examinations. This process often delays treatment until substantial dopaminergic loss has already occurred. In this paper we describe a multiagent Retrieval-Augmented Generation (RAG) platform that automates PD stage prediction directly from T1-weighted MRI. The pipeline runs on edge hardware under a hybrid cloud arrangement: an EfficientNet-B0 clas-sifier performs on-device inference while three cooperating agents (a Supervisor, an AI/ML module, and a RAG module) coordinate through shared memory to generate citation-backed clinical reports. Training and evaluation used the PPMI cohort (39,534 slices drawn from 1,064 subjects). On this data the classifier reaches 99.62% accuracy at the scan level and 91.4% under subject-level five-fold cross-validation, spanning six categories (healthy con-trol plus Stages 1–5). Coupling the classifier with the RAG agent cuts large-language-model hallucination from 18.5% down to 2.0% and lifts citation accuracy from 12% to 96%. A component-wise ablation confirms that each agent makes a distinct, measurable contribu-tion. Grad-CAM heatmaps, validated against atlas-defined neuroanatomical regions, show statistically significant overlap with structures known to be affected in PD. Finally, a 100-case de-identified workflow simulation yielded a 40% reduction in documentation time together with a clinician–system agreement of Cohen’s \(\bf\kappa\) = 0.92.