Abstract Body (Do not enter title and authors here): Background: Fewer than half of eligible patients with heart failure (HF) receive all 4 guideline-directed medical therapies (GDMT), despite strong evidence of benefit. Traditional electronic monitoring of GDMT adoption relies on structured EHR data, which fails to capture barriers documented in clinical notes.

Hypothesis: Retrieval-Augmented Generation (RAG), which uses large language models (LLMs) to parse both structured and unstructured EHR data, can significantly improve the identification of actual GDMT use/underuse compared with structured data alone.

Methods: We developed a RAG system that parses tabular data and clinical notes to identify barriers to GDMT utilization. The system employed a 2-stage retrieval approach, combining semantic search with relevance re-ranking and temporal EHR analysis that spanned 1 year before and after hospitalization. Detection rates of reasons for non-utilization were compared between structured-only and RAG-based approaches using chi-square tests and evaluated in an expert review.

Results: Among 26,175 HF patients in the Yale New Haven Health System between 2013 and 2024 (mean age 82.4±15.0 years, 49.1% female), RAG significantly improved the identification of barriers to GDMT across all drug classes. For beta-blockers, RAG increased the detection of bradycardia by 25.8% (from 89 to 112 cases) and hypotension by 31.9% (from 72 to 95 cases), uncovering 31 previously undocumented cases of asthma/COPD and 21 patient refusals. For ACE/ARB/ARNI, RAG improved identification of hyperkalemia by 25.4% (67 to 84 cases) and renal dysfunction by 21.4% (103 to 125 cases), while also identifying 41 additional cases of cough, 27 of angioedema, and 7 pregnancy concerns. Barriers to MRA use included a 26.2% increase in hyperkalemia detection (from 84 to 106 cases) and of gynecomastia (19 cases). SGLT2i had the largest gaps in documentation, with 67 cases of UTI missed in structured data. RAG also revealed systematic barriers not captured in structured fields, including cost/insurance-related issues (178 SGLT2, 52 ACE/ARB/ARNI, 21 beta-blocker cases).

Conclusion: RAG improves the identification of reasons for GDMT underutilization by extracting insights from unstructured EHR data, outperforming information retrieved from structured data alone. The automated detection of care gaps can enable targeted interventions, with the potential for scalable, interoperable deployment across health systems.