MedInsight Pro: An Explainable Hybrid NLP and LLM Fusion Framework for Radiology Report Interpretation

Abstract

Radiology reports are often unstructured, containing ambiguous phrases and negations that make automated interpretation difficult. This paper introduces MedInsight Pro, a hybrid clinical NLP (Natural Language Processing) framework that fuses deterministic parsing with LLM (Large Language Model) augmentation to extract meaningful insights from chest X-ray impressions. The system combines rule-based negation detection, dependency parsing, and lexicon-driven keyword matching with contextual expansion powered by Groq’s LLM. On an expanded CheXpert-derived cohort of 2 020 reports (1 020 real and 1 000 augmented samples), MedInsight Pro achieved an average macro F1-score of 89.7%, with overall precision of 91.8%, recall of 89.4%, and negation detection accuracy of 95.1%. These results surpass keyword-only, LLM-only, and CheXpert labeler baselines. Deployed via a Streamlit interface, the system also offers explainability through clause heatmaps, radar charts, robustness visualizations under synthetic domain shift, and a clinical Q&A interface. The findings highlight the practicality of hybrid NLP–LLM pipelines as scalable, transparent decision-support tools for radiologists in real-world clinical settings. The system demonstrates strong performance in structured label extraction and offers interpretability through visualization modules and an LLM-assisted reasoning layer.