Resp-GRASP: Clinically-Grounded Reasoning for Respiratory Signal Interpretation via Guardrailed RAG with Personalized Baselines

Although continuous monitoring allows early identification of physiological deterioration five days prior to actual symptomatic COPD events, there exist two major limitations: predictions in a “black box” format which cannot be reviewed by clinicians, and population-based threshold settings resulting in false positives for abnormal measurements. This paper introduces Resp-GRASP, a six-step guardrailed reasoning pipeline that solves both limitations using double-layer z-score hierarchy that favors personal trends over population statistics, as well as MedGemma-27B that generates clinical reports while citing evidence-based ATS/ERS guidelines. Resp-GRASP achieves selective prediction in a multi-step setting, aborting 18.9% of recordings with low signal discriminability to ensure reliable inference. The primary finding of this study is the six-fold reduction in false positive rates (15.8% → 2.6%) while nearly doubling sensitivity (53.3% → 93.3%) with McNemar test p = 0.031 in comparison with approaches based on population thresholds. We introduce Semantic Directional CRC (SD-CRC), a step-level reasoning verification metric that checks whether each step’s clinical narrative directionally agrees with its quantitative z-score evidence after stripping retrieved passages to prevent evidence contamination from inflating scores. Resp-GRASP has an accuracy of 85.7% SD-CRC (p < 0.0001).