Back

RAGnosis: Retrieval-Augmented Generation for Enhanced Medical Decision Making

Rouhollahi, A.; Homaei, A.; Sahu, A.; Harari, R.; Nezami, F. R.

2025-06-12 health informatics
10.1101/2025.06.11.25329438 medRxiv
Show abstract

We present RAGnosis, a fully offline, retrieval-augmented framework for interpreting unstructured clinical text using open-weight large language models. As a proof of concept, we apply RAGnosis to the task of paravalvular leak (PVL) classification from cardiac catheterization reports, a process that typically requires slow, expert-driven interpretation. The system combines local OCR, semantic retrieval, and instruction-tuned LLMs to generate evidence-backed predictions and explanations grounded in real clinical documentation. We evaluate four models (DeepSeek 70B, Gemma 27B, Mistral 7B, LLaMA 3B) across 100 reports, analyzing classification accuracy, explanation quality, and retrieval relevance. Results highlight tradeoffs between fluency and reliability, with DeepSeek demonstrating the most consistent performance. By operating entirely on-prem and supporting modular integration, RAGnosis provides a scalable and interpretable foundation for clinical NLP that delivers not just answers but traceable reasoning.

Matching journals

The top 2 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.