Heart teams and AI assistants: practical lessons from real-world implementation of clinical guideline support

Heart team decision-making involves complex guideline interpretation under time pressure, yet clinical AI implementation often fails due to poor understanding of human-AI interaction dynamics and inadequate workflow integration.

To evaluate practical implementation and functional capabilities of an iteratively refined AI-powered clinical guideline assistant in real-world cardiology practice, focusing on deployment challenges and specialist heart team interaction patterns.

Over 18 months, we iteratively developed a guideline-focused AI assistant through three phases: custom language model proof-of-concept, cloud infrastructure exploration, and enterprise SaaS implementation. Early phases (N=412 interactions with multidisciplinary specialist heart team including cardiologists, fellows, and specialist nurses) informed design refinements addressing usability, cost-effectiveness, and guideline adherence challenges. This analysis focuses on 120 anonymized clinical interactions (average 5.52 messages/conversation) via web interface and messaging platform during final deployment phase, characterizing functional capabilities and user engagement patterns. The AI provided ESC guideline-based decision support across tertiary and district hospital settings.

Iterative development through earlier phases (N=412 interactions) informed design refinements that enhanced user engagement and platform usability in the final enterprise SaaS deployment. Final phase analysis (N=120 interactions) revealed sophisticated clinical decision support capabilities across diverse cardiovascular domains including acute coronary syndromes, heart failure, arrhythmias, and cardiomyopathies. Frequent utilization patterns included ESC guideline-specific information retrieval, dynamic risk score calculations (RF-CL, CHA₂DS₂-VASc, H2FPEF), ECG interpretation guidance, and complex pharmacological decisions. User interaction analysis demonstrated preference for conversational, iterative, scenario-based queries with evidence of clinical reasoning refinement through multi-turn conversations. International usage across multiple countries confirmed broader clinical applicability. Key success factors included continuous clinician-led development, seamless workflow integration, multilingual capability, and clear AI scope boundaries.

Successful clinical AI deployment depends more on human-AI interaction design and workflow integration than algorithmic sophistication. Clinician-centered iterative development yields measurable improvements in user engagement and platform usability. Our analysis confirms AI capability to provide nuanced, guideline-conformant decision support across diverse cardiological scenarios, demonstrating significant potential for clinical decision support and medical education in cardiovascular practice through evidence-based human-AI collaboration.

Questions word cloud