ARCHIVES

Traditional Clinical Decision Support Systems (CDSS) were built on rule-based engines and standard-alone machine learning classifiers. The emergence of Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) upgraded these systems by enabling knowledge-grounded reasoning and dynamic biomedical evidence retrieval. Recent multi-agent LLM frameworks further integrated biomedical RAG with external web intelligence. However, most existing approaches rely on heuristic score aggregation and lack explicit conflict modeling and probability calibration, limiting reliability in safety critical healthcare environments. The system proposes a conflict aware, machine-learned, and probability-calibrated multi-evidence fusion framework that transforms heuristic reasoning into a structured probabilistic decision model. Biomedical RAG outputs and external web evidence are fused using a supervised ensemble model that learns nonlinear interactions between evidence signals. Engineered features capture diagnostic agreement and disagreement with model reliability. Isotonic regression is applied for probability calibration and evaluated using Brier Score and Expected Calibration Error (ECE). SHAP-based feature attribution ensures transparent and interpretable diagnostic confidence decisions, improving robustness and clinical trust in LLM-based CDSS