HEAL-RAG: A Hybrid Evolving Retrieval-Augmented Generation Framework for Explainable Biomedical Data Analysis and Clinical Decision Support

The rapid growth of biomedical literature, electronic health records (EHRs), and heterogeneous clinical data has created a need for systems that can synthesise evidence to support clinical decision-making. Large language models (LLMs) reason fluently in natural language, but their clinical use is limited by hallucination, gaps in domain knowledge, opaque reasoning, and an inability to incorporate knowledge published after training. We present HEAL-RAG (Hybrid Evolving Architecture for LLM-enhanced Retrieval-Augmented Generation), a framework that combines four components in one pipeline: (1) a hierarchical multi-vector retriever that queries structured and unstructured biomedical knowledge sources (PubMed, UMLS and clinical guidelines) with hierarchy- and recency-aware ranking; (2) a multimodal fusion module that integrates clinical narratives, laboratory results, structured radiology reports, and genomic findings into a shared representation; (3) a memory-based feedback component with episodic and semantic memories; and (4) a provenance-aware explanation module that links each generated claim to its supporting sources. The individual components build on prior work; our contribution is their integration and a component-level evaluation. We evaluated HEAL-RAG on three public text benchmarks: biomedical question answering (PubMedQA), rare-disease diagnostic reasoning from the clinical histories and radiologist captions of the NOVA brain-MRI benchmark, and pharmacogenomic guideline questions (PGxQA). On PubMedQA, HEAL-RAG achieved an accuracy of 84.2% and an F1 score of 0.876, which is 5.8 percentage points above the same LLM without retrieval (LLaMA-3.1-70B) and 1.9 percentage points above the strongest reproduced baseline (hierarchical RAG). Among the retrieval-based systems, HEAL-RAG had the highest proportion of factual claims supported by the cited evidence (98.5%, compared with 76.3% for standard RAG). In an ablation study, removing any single component other than the explanation module reduced accuracy. These are benchmark results only: the multimodal and feedback components were not evaluated in a clinical or longitudinal setting, and prospective clinical validation is required before any clinical use.