Reply

We read with interest the letter by Dr Bareeqa and colleagues. The criticisms raised regarding our paper are addressed below.

First, concerns regarding the absence of technical architecture, validation workflow, and safety framework are unfounded and can be found in the JACC: Advances Methods Companion,¹ which describes these elements comprehensively. That paper explicitly details the system architecture (retrieval-augmented generation [RAG]-based, guideline-anchored large language model), the multistage validation process based on real-world cases with iterative expert refinement, and the embedded safety logic with red flags and physician oversight. Second, the claim of inadequate predeployment validation is inaccurate. The Methods Companion outlines a structured validation process involving hundreds of real-world cases, comparison with cardiologist judgment, and iterative refinements to ensure guideline-aligned performance, demonstrating rigorous preclinical evaluation.

Third, the critique of RAG is generic and not applicable to the system under study. SIRIO-HF is a constrained, guideline-grounded clinical decision support system, with structured inputs and safeguards, so applying limitations of unconstrained RAG systems reflects a misunderstanding of the architecture.

Fourth, the authors have misinterpreted the trial. ASSIST-HF SIRIO did not evaluate “AI vs clinicians,” but rather a supervised decision support tool with mandatory human-in-the-loop governance, in which every recommendation generated by the virtual assistant (VA) was reviewed, modified if needed, and ultimately authorized by a cardiologist, who retained full decision-making authority.

Fifth, the link between AI use and increased visit frequency is misinterpreted. More frequent contact, enabled by the AI model to address specialist capacity limits, was an intentional part of the care design, not a confounder. The study evaluated a scalable, supervised workflow combining frequent monitoring and AI support, rather than isolating AI’s independent effect.

Finally, the statement regarding “zero disagreements” is wrong. The JACC Brief Report² clearly states that cardiologist–VA agreement on guideline-directed medical therapy optimization was 100% in the early phase of the trial and 93.4% at the end of treatment. Disagreements were few and mainly reflected perceived “overcaution” in VA recommendation, such as extreme caution in the presence of borderline potassium values. The Methods Companion paper¹ explicitly states that all VA-clinician disagreements were prospectively logged and reviewed within a structured governance framework, with continuous monitoring and embedded disagreement analysis for safety.

The concerns of Bareeqa et al can be addressed by a complete consideration of the Methods Companion and an understanding of the supervised, workflow-based intervention. ASSIST-HF SIRIO does not assess autonomous AI, but a physician-supervised model designed to reduce specialist input, enable more frequent visits, and improve scalability and standardization of care.