An artificial intelligence (AI) virtual assistant supervised by nonclinical staff helped recently hospitalized patients with new diagnosis of heart failure with reduced ejection fraction (HFrEF) achieve guideline-directed medical therapy optimization, a new pilot study suggested.

In the study, published online in the Journal of the American College of Cardiology, the results were achieved more rapidly and more often than patients who received usual care from a cardiologist or an HF nurse.

The workflow involved a virtual generative AI assistant built with a large language model that provided real-time treatment recommendations to a nonclinical administrator and a remote cardiologist to sign off on all recommendations. The results showed patient-reported acceptability and appropriateness of the virtual assistant care workflow was high.

Although the clinical outcomes were exploratory due to the study’s small sample size, the results point to a potential solution for an unmet medical need: the application and scaling of timely guideline-directed medical therapy (GDMT) for the hospitalized patients with HFrEF.

“Patients discharged after HFrEF hospitalization remain at high early risk, yet GDMT initiation and rapid uptitration are often delayed because frequent follow-up is difficult to deliver at scale,” first author Eliano P. Navarese, MD, PhD, of the Department of Life and Health Sciences at Link Campus University in Rome, Italy, told Medscape Medical News.

“We designed this study to address therapeutic inertia by increasing the cadence and consistency of medication optimization while preserving cardiologist accountability for every decision,” he said.

Virtual Assistant-Guided Care

The ASSIST HF SIRIO pilot study included 60 patients at a single center who were randomized to receive usual care with either a cardiologist or HF nurse at regular clinic visits, or care every 2 weeks from the AI workflow. Researchers compared the feasibility, acceptability, and rate of GDMT optimization in each group at 12 weeks.

In the AI group, the administrator used a scripted interview to gather data on symptoms, vital signs, weight, and laboratory values. Those results were entered into the virtual assistant, which then generated a personalized treatment recommendation. Next the remote cardiologist approved, modified, or rejected the recommendation before it was sent to the patient.

“GDMT optimization is a structured, guideline-driven process that depends on serial clinical inputs. A large language model can support this only if used within a constrained, safety-first architecture — not as an open-ended chatbot,” Navarese said. “We used retrieval-augmented generation to ground recommendations in curated guideline content, combined with expert prompt engineering to enforce safety checks and reduce hallucinations.”

High Marks From Patients

Patients who received care guided by a virtual assistant reported high levels of acceptance (90%), appropriateness (92%), and feasibility (88%) for the model. And 87% of patients reported a greater than 80% score on a combined acceptability measure.

Patients in the virtual assistant group achieved a higher rate of maximally tolerated doses of GDMT at 12 weeks than those in the usual care group. Results included patients who received beta blockers (100% vs 52%; P < .001), angiotensin receptor/neprilysin inhibitors (100% vs 37%; P < .001), angiotensin-converting enzyme inhibitors and angiotensin II receptor blockers (100% vs 35%; P < .001), and mineralocorticoid receptor antagonists (96% vs 20%; P < .001) as well as SGLT2 inhibitors (96% vs 20%; P < .001).

Navarese said the virtual assistant model “can increase the cadence of GDMT optimization without replacing physician judgment.”

“The key lesson is implementation, governance, and safety-by-design: Generative AI can be deployed responsibly when embedded in a supervised clinical workflow with auditable outputs and explicit accountability,” he said.

While AI technology in cardiology is usually focused on prediction models or risk scores, “this work is different,” Navarese said.

“[I]t places generative AI inside a real therapeutic pathway under explicit governance —guideline grounding, constraint-based generation, audit trails, red-flag escalation, and mandatory physician oversight,” he said.

An Unmet Need

The study is “a very important proof of concept pilot study that addresses an unmet need: How to provide intensive medication adjustment in the context of limited HF specialist bandwidth,” Trejeeve Martyn, MD, MSc, staff cardiologist in the Section of Heart Failure and Cardiac Transplantation at the Cleveland Clinic, Cleveland, told Medscape Medical News.

“We have nearly 7 million patients with heart failure living in the United States and given epidemiologic trends in obesity and an aging population, heart failure prevalence is projected to increase in the coming years,” said Martyn, who is also director of Heart Failure Population Health for Cleveland Clinic Health System.

Between 25% and 30% of US patients receive cardiovascular follow-up after hospitalization for HF, and about half receive follow-up from a cardiologist in the 30 days after discharge, Martyn noted.

“We have a massive access and capacity problem in the heart failure space when it comes to transitions of care. Therefore, automated approaches to optimizing medical management are going to be impactful,” he said.

The study is “a great example of AI providing algorithmic clinical decision support, which may reduce the need for scarce specialized cardiovascular resources,” Martyn said.

“The virtual assistant clearly showed a benefit in titration of these medicines without adverse outcomes, and that’s the main takeaway for all of this,” he said.

Future studies could examine how cardiologist oversight is best applied to scale the virtual assistant model, evaluated workflow and staffing-related questions, and analyze how patient symptom information is transmitted electronically, Martyn said.

“Lastly, patient acceptance, attitudes, and willingness to take part in this kind of care, along with how best to set expectations should be evaluated with the emergence of agentic AI with clinicians-in-the-loop as the likely future of chronic HF management,” he said.

The ASSIST-HF SIRIO trial was supported by the East and North Hertfordshire NHS Trust. Navarese reported having no relevant financial disclosures. Martyn reported being an ad hoc reviewer for the Journal of the American College of Cardiology and has been a consultant to AstraZeneca and Bayer for heart failure medical therapy-like agents related to GDMT.

Jeff Craven is an independent journalist living in Wilmington, Delaware.