Authors: Assistant Professor A.Saranya, Assistant Professor Dr. Swarna Surekha
Abstract: The use of large language models (LLMs) in CDSS promises groundbreaking advances but faces challenges owing to the "black box" nature of these models. MedGPT-XAI is proposed in this paper as an innovative framework combining domain-adapted LLMs with multimodal explainability for CDSS. MedGPT-XAI consists of three modules, namely BioBERT, fine-tuned GPT-2 and an explainability module containing ALTI-Logit Decomposition, Attention Manipulation, Evidence-Graph Verification and LightGBM Feature Importance. Experiments conducted on MedQA and MIMIC-III show 89.8% accuracy in decision making with 120ms response time while the metrics for explaining reasoning chains yield 78% evidence support rate and 72% reasoning coherence. Comparison between this framework and baseline models show 15-20% gains in clinician trust scores, thus establishing the viability of MedGPT-XAI framework for clinical applications of AI.
