ClinicGraphRAG knowledge graph and LLM retrieval-augmented generation for reliable clinical decision support summaries

Large language models can generate fluent clinical summaries, yet their reliability is constrained by hallucinations, limited provenance, and safety risks when used in conjunction with automated image-based predictions. This study presents an end-to-end screening-oriented decision support pipeline that integrates a dual-path vision model with a domain-constrained retrieval-augmented generation framework to produce evidence-grounded explanations. The vision component, Attention-Guided Dual-Path FusionNet, combines a ConvNeXtV2 backbone for global semantic representation with a Swin Transformer branch for hierarchical local feature modeling, followed by attention-guided feature fusion for four-class CT image classification. The language component, ClinicGraphRAG, performs allowlisted retrieval over curated clinical sources with metadata-preserving provenance and generates constrained outputs using a grounded-only input interface. The dataset comprises 1,535 CT images expanded to 3,600 samples through class-specific, training-only augmentation under a leakage-safe stratified 5-fold cross-validation protocol, which may introduce bias due to reliance on synthetic data and limited sample diversity. The proposed model achieves a mean cross-validation accuracy of 98.74 ± 0.19, outperforming selected baseline architectures; however, these results are derived from a single curated dataset and may not generalize to heterogeneous clinical settings. The system is designed strictly for screening-oriented decision support and education, with explicit constraints to prevent diagnostic or treatment use, and requires external validation and broader cohort evaluation before real-world deployment.