An Edge-Deployed Explainable AI Framework with RAG-Augmented Large Language Models for Multi-Stage Alzheimer’s Disease Staging and Conversational Decision Support

Main Article Content

Saksham Srivastava
Nishita Singh
Dr. Rishi jain
Mr. Aakash Kuragayala
HIMADRI SRIVASTAVA
SAUD AHMAD KHAN

Abstract

Alzheimer’s disease (AD) demands precise multi-stage classification and interpretable diagnostic workflows to support timely clinical intervention and personalised patient management. While deep learning models have shown strong predictive capabilities, conventional vision architectures frequently operate as uninterpretable "black boxes," failing to highlight neuroanatomical regions of interest or provide contextually grounded, actionable clinical guidance for clinicians, patients, and caregivers. To address these critical gaps, this paper presents a novel, edge-integrated preliminary decision-support framework that unifies deep convolutional neural networks, explainable artificial intelligence (XAI), and retrieval-augmented conversational intelligence for end-to-end Alzheimer’s disease assistance. The computer vision pipeline utilises a fine-tuned VGG16 architecture trained to classify structural magnetic resonance imaging (MRI) scans across distinct stages of AD progression, ranging from non-demented to moderate cognitive impairment. To eliminate diagnostic opacity, Gradient-Weighted Class Activation Mapping (Grad-CAM) is embedded into the neural network pipeline to generate visual heatmaps, precisely isolating affected focal regions and localised neuropathological atrophy patterns within brain tissue. Beyond visual staging, the system incorporates an edge-deployed Llama 3.2 Large Language Model (LLM) integrated with a Retrieval-Augmented Generation (RAG) framework. This local RAG pipeline is domain-grounded using clinical guidelines established by the Alzheimer’s Association, enabling privacy-preserving, low-latency execution directly on local edge hardware without reliance on external cloud services or compromising sensitive patient health data. Experimental evaluation demonstrates that the proposed VGG16 classifier achieves a multi-class staging accuracy of 90.0%, showing robust diagnostic feature extraction across severity levels. The Grad-CAM visual explanations effectively corroborate automated predictions with established neuroimaging biomarkers, significantly enhancing model interpretability and clinician trust. Concurrently, the integrated Llama 3.2 RAG conversational agent provides a multi-functional user interface, enabling real-time interactive query resolution and automated generation of structured, evidence-grounded diagnostic reports. By synergising deep learning classification, visual XAI, localised LLM reasoning, and domain-specific knowledge bases, this comprehensive edge AI architecture offers an interpretable, secure, and highly accessible preliminary diagnostic tool for modern neurodegenerative disease management.

Article Details

Section

Articles

How to Cite

[1]
“An Edge-Deployed Explainable AI Framework with RAG-Augmented Large Language Models for Multi-Stage Alzheimer’s Disease Staging and Conversational Decision Support”, DTMST, vol. 1, no. 1, Nov. 2026, doi: 10.52939/b6cvw238.