A Natural Language Processing–Driven Electronic Medical Record Program to Optimize Medical Therapy for Heart Failure in Community Practices

Background Heart failure with reduced ejection fraction (HFrEF) carries high morbidity and mortality. Guideline-directed medical therapy (GDMT) improves outcomes, yet implementation in routine practice remains inadequate. Community cardiology practices in Canada show suboptimal adherence to GDMT for HFrEF, highlighting a gap between evidence and care. Methods We conducted a prospective, two-phase, before-and-after observational quality-improvement study (the Heart Failure iDentification and Optimization in the Community [HFiDOC] program) in Canadian community cardiology practices (Phase 1, Ontario, 2021–2022; Phase 2, pan-Canadian, 2022–2024). Patient identification used EMR cell queries and natural language processing (NLP; DARWEN™ and Apollo™). In Phase 1, cardiologists received aggregate practice-level reports of GDMT use (traditional audit-and-feedback); in Phase 2, NLP-derived, patient-specific recommendations prioritized individuals for GDMT initiation or intensification delivered via secure dashboards or electronic fax. The primary outcome was the change in GDMT prescription rates. Results A total of 1,796 HFrEF patients followed by 57 community cardiologists were identified. At baseline, fewer than half received an MRA or SGLT2i, and only 22.7% were treated with all four GDMT classes. After a median 259 days, the proportion on quadruple therapy increased from 15.4% to 74.9% (p