CardiacGPT™: A Real-Time AI Assistant for Intraoperative Guidance and Postoperative Decision Support in Cardiac Surgery

Abstract

Abstract Background Cardiac surgery is one of the most complex and high-stakes areas of medicine, where intraoperative decisions must be made within seconds and incomplete information can compromise outcomes. Traditional risk scores and rule-based decision support tools provide limited real-time guidance and rarely integrate the unstructured data streams available during surgery. Recent advances in large language models (LLMs) such as OpenAI’s GPT-5 and Anthropic’s Claude 3.5 family have demonstrated state-of-the-art reasoning, summarization, and clinical dialogue capabilities. However, their safety and trustworthiness in surgical settings remain untested. Objective To evaluate the feasibility and clinician trustworthiness of CardiacGPT , a real-time AI assistant that leverages the newest-generation LLMs for intraoperative guidance and postoperative decision support. Methods We retrospectively analyzed 500 de-identified cardiac surgery cases from Brigham and Women’s Hospital, including CABG, valve, and combined procedures. Structured EHR variables, intraoperative monitoring, and operative notes were formatted into standardized prompts and processed through four cutting-edge models: OpenAI GPT-5, Anthropic Claude 3.5 Opus, Claude 3.5 Sonnet, and Claude 3.5 Haiku. Outputs were presented via a blinded Bidding App to attending cardiac surgeons and ICU clinicians, who scored trust and clinical relevance on a 5-point Likert scale. The primary outcome was the proportion of high-trust ratings (score > 4); secondary outcomes included mean trust scores, variance, and inter-rater reliability. Results Across 2,000 evaluations, GPT-5 and Claude 3.5 Opus achieved the highest mean trust scores (4.83 and 4.79, respectively), each exceeding 98% high-trust ratings. Claude 3.5 Sonnet performed moderately (mean 3.9, 74% high-trust), while Claude 3.5 Haiku produced less context-specific recommendations (mean 3.6, 66% high-trust). Inter-rater reliability was excellent, with ICC(2,1) = 0.91 (95% CI 0.88–0.94), confirming strong agreement among reviewers. Qualitative analysis showed that GPT-5 and Claude 3.5 Opus generated actionable and context-aware outputs, whereas smaller models often produced generic or incomplete guidance. Conclusions CardiacGPT, powered by the newest LLMs (GPT-5 and Claude 3.5 series), demonstrated feasibility and exceptionally high clinician trust across 500 real-world surgical cases. This is the first blinded, multi-model evaluation of next-generation LLMs for cardiac surgery. While outcome-based prospective trials are still required, these results establish CardiacGPT as a promising real-time co-pilot for cardiac surgeons, with the potential to reduce cognitive load, standardize intraoperative communication, and improve postoperative planning.

Ähnliche Arbeiten