MMCRAG-Resp: a multi-modal corrective retrieval-augmented generation framework for explainable respiratory disease...

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Respiratory Medicine

Background:

• Standard Retrieval-Augmented Generation (RAG) systems rely solely on semantic similarity, often retrieving clinically irrelevant evidence, leading to unsafe or hallucinated outputs in critical fields like respiratory care.
• Accurate diagnosis in respiratory medicine heavily depends on precise physiological indicators, including spirometry patterns and symptom profiles, which current RAG systems struggle to interpret effectively.

Purpose of the Study:

• To introduce MMCRAG-Resp, a novel framework for respiratory intelligence that enhances clinical decision support through physiology-aware corrective retrieval.
• To address the limitations of standard RAG by incorporating clinical grounding and physiological awareness into the evidence retrieval process for respiratory care.

Main Methods:

• Harmonized 17,516 patient records from diverse datasets (Respiratory Sound Database, NHANES spirometry, clinical guidelines) using a unified respiratory ontology.
• Developed multi-modal embeddings (acoustic, spirometric, textual) fused into 144-dimensional vectors and indexed in FAISS for efficient retrieval.
• Implemented a Respiratory Relevance Score (RRS) combining spirometry pattern agreement, symptom overlap, and embedding similarity to gate evidence quality, followed by a three-path correction policy for LLM consumption using a locally deployed small language model.

Main Results:

• MMCRAG-Resp achieved a Clinical Precision@10 of 0.81 and a Spirometry Alignment Score of 0.79 on 200 held-out test queries.
• Demonstrated a significant reduction in hallucination rate (0.11, 71% relative reduction compared to Vanilla RAG) and a high Evidence Citation Rate of 0.87.
• Ablation studies confirmed the critical contribution of spirometry-first scoring, with statistically significant improvements over baselines (p < 0.001, Cohen's d = 1.42).

Conclusions:

• MMCRAG-Resp represents a pioneering system integrating multi-modal respiratory evidence retrieval, physiologically-grounded corrective RAG, and privacy-preserving local LLM inference.
• The framework advances safe, explainable, and privacy-preserving AI-assisted respiratory decision support, offering a deployable clinical reasoning solution.