Large language model recommendations for heart failure therapy

Background Guideline-directed medical therapy (GDMT) for heart failure (HF) reduces mortality and hospitalizations but remains underutilized, in part because contraindications, laboratory trends, and prior medication intolerances are time-consuming for clinicians to synthesize. Large language models (LLMs) may help integrate structured and unstructured EHR data to support GDMT optimization. Purpose To develop and validate an LLM-based workflow for patient-specific GDMT suggestions. Methods Consecutive patients with HF scheduled for outpatient cardiology follow-up visits at our hospital, from January–March 2026 were identified. Patients with restrictive/hypertrophic cardiomyopathy, severe left-sided valve disease, or prior transplant/LVAD were excluded. Each patient’s recent vital signs, medications, key laboratory results, echo reports, and clinician notes were extracted from the EHR using an automated query. These data were processed through a purpose-built workflow powered by a secure OpenAI GPT-5.2 model, to confirm HF diagnosis and eligibility, and to generate patient-specific, evidence-based, provider-facing suggestion messages. Each message included a summary of relevant patient data, a list of GDMT suggestions, and suggestion-specific rationales and safety checks (Figure 1). Cardiologists evaluated the safety of 200 randomly selected messages with ≥1 suggestion. Results Among 1,285 scheduled patients, 1,016 were eligible for decision support and 604 had ≥1 GDMT suggestion. Among patients with suggestions, median age was 73 [62-80] years, 265 (44%) were female, and median ejection fraction was 48% [36%-60%]. Baseline GDMT utilization included beta blockers in 521 (87%), ARNI in 156 (26%), MRAs in 230 (38%), and SGLT2i in 267 (44%). The GDMT suggestions were non-mutually exclusive and included initiation or up-titration of SGLT2i in 297 patients (49%), non-steroidal MRA in 177 (29%), ARNI in 168 (28%), steroidal MRA in 158 (26%), beta blocker in 69 (11%), and ARB in 10 (2%), with expected variation by ejection fraction (Figure 2). In the safety sample, 6/200 messages (3%) were initially judged as potentially unsafe by the cardiologist reviewer, of which 2 were considered potentially unsafe by consensus of a group of cardiologists. Reasons for the 2 unsafe messages were recommending SGLT2i in one patient with creatinine 3.1 mg/dL and another with severe frailty, malnutrition, and previous intolerance to GDMT. Conclusions A purpose-built LLM workflow generated patient-specific, provider-facing GDMT suggestion messages that were 99 % safe on consensus review. A randomized trial will evaluate whether communicating these messages to providers before outpatient visits improves HF GDMT utilization.Figure 1Figure 2