AI-Automated Risk Operative Stratification for Severe Aortic Stenosis: A Proof-of-Concept Study.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Accurate operative risk stratification is crucial for treatment selection in severe aortic stenosis.
• Multidisciplinary Heart Teams provide essential risk assessment for these patients.
• Automated workflows using large language models (LLMs) were explored to replicate Heart Team assessments.

Purpose of the Study:

• To develop and evaluate an automated LLM-based workflow for replicating Heart Team risk stratification in severe aortic stenosis.
• To compare the accuracy of LLM-based risk calculations against traditional methods and Heart Team decisions.

Main Methods:

• A retrospective analysis of 231 severe aortic stenosis patients was conducted.
• An automated GPT-4o system was developed for data extraction, EuroSCORE II calculation (algorithmic vs. LLM), and risk stratification.
• Risk stratification accuracy was assessed by comparing LLM approaches (with/without EuroSCORE II) against Heart Team decisions.

Main Results:

• The LLM workflow processed patients rapidly (32.6 ± 6.4 seconds).
• LLM-calculated EuroSCORE II showed a smaller mean difference from Heart Team values compared to algorithmic calculation.
• A guideline-integrated LLM approach without EuroSCORE II achieved the highest accuracy (90.0%) and AUC (0.93) for risk stratification.
• Systematic hallucinations were observed for cardiovascular risk factors when data were missing.

Conclusions:

• LLMs demonstrated accuracy in EuroSCORE II calculation and achieved high concordance (90%) with Heart Team decisions.
• Despite promising accuracy, LLM hallucinations, reproducibility issues, and lack of clinical outcome validation prevent immediate clinical application.
• Further research is needed to address LLM limitations before clinical integration in severe aortic stenosis management.