Comprehensive Summary
This study analyzes the congruence between clinical recommendations for coronary artery disease (CAD) between that of an artificial intelligence language model (ChatGPT) versus a heart team. Specifically, they were interested in comparing the treatment strategies of percutaneous coronary intervention (PCI) and coronary artery bypass grafting (CABG). The research team created a standardized set of clinical vignettes for 137 patients and submitted them to ChatGPT. They then compared the AI’s recommendations of PCI vs CABG against the heart team’s decisions. The AI model had an overall concordance rate of 65% with the HT, showing stronger agreement when the heart team chose CABG (82.4%) but weaker agreement when the heart team chose PCI (44.4%). Discordance was most common in older patients, those with diabetes, and those with impaired renal function, with AI tending to over recommend CABG relative to the heart team. The authors emphasize that while AI could somewhat reliably generate recommendations, it lacked certain judgment needed for PCI cases as compared to the trained heart team.
Outcomes and Implications
Multivessel coronary artery disease is common and carries significant morbidity and mortality, and treatment decisions between CABG and PCI have major consequences for patient outcomes. This research is important because it explores whether AI could augment or streamline complex decision-making processes traditionally requiring a multidisciplinary team. Clinically, the findings suggest that current AI models may be useful for decision support—particularly in identifying strong CABG candidates—but they are not yet accurate enough to replace the Heart Team’s expertise, especially when PCI may be appropriate. Since the model used was a general-purpose LLM and not trained on multimodal imaging or cardiology-specific data, substantial refinement and validation are needed before integration into routine practice. The authors note AI could eventually help reduce workload and improve access to guideline-consistent care, but widespread clinical implementation will require specialized, multimodal models and rigorous testing in prospective trials.