Exploring the Potential of ChatGPT-4 for Clinical Decision Support in Cardiac Electrophysiology and Its Semi-Automatic Evaluation Metrics

Background/Aim: Despite extensive research in other medical fields, the capabilities of ChatGPT-4 in clinical decision support within cardiac electrophysiology (EP) remain largely unexplored. This study aims to enhance ChatGPT- 4`s domain-specific expertise by employing the Retrieval-Augmented Generation (RAG) approach, which integrates up-to-date, evidence-based knowledge into ChatGPT-4`s foundational database. Additionally, we plan to explore the use of commonly used automatic evaluation metrics in natural language processing, such as BERTScore, BLEURT, and cosine similarity, alongside human evaluation, to develop a semi-automatic framework. This aims to reduce dependency on exhaustive human evaluations, addressing the need for efficient and scalable assessment tools in medical decision-making, given the rapid adoption of ChatGPT-4 by the public. Method: We analyzed five atrial fibrillation (Afib) cases and seven cardiac implantable electronic device (CIED) infection cases curated from PubMed case reports. We conducted a total of 120 experiments for Afib and 168 for CIED cases, testing each case across four temperature settings (0, 0.5, 1, 1.2) and three seed settings (1, 2, 3). ChatGPT-4`s performance was assessed under two modes: the Retrieval-Augmented Generation (RAG) mode and the Cold Turkey mode, which queries ChatGPT without external knowledge via RAG. For Afib cases, ChatGPT was asked to determine rate, rhythm, and anticoagulation options, and provide reasoning for each. For CIED cases, ChatGPT is asked to determine the presence of device infections. Accuracy metrics evaluated the determination component, while reasoning was assessed by human evaluation, BERTScore, BLEURT, and cosine similarity. A mixed effects analysis was used to compare the performance under both models across varying seeds and temperatures. Spearman`s rank correlation was used to explore the relationship between automatic metrics and human evaluation. Results: In this study, 120 experiments for Afib and 168 for CIED were conducted. There is no significant difference between the RAG mode and the Cold Turkey mode across various metrics including determination accuracy, reasoning similarity, and human evaluation scores, although RAG achieved higher cosine similarity scores in Afib cases (0.82 vs. 0.75) and better accuracy in CIED cases (0.70 vs. 0.66), though these differences were not statistically significant due to the small sample size. Our mixed effects analysis revealed no significant effects of temperature or method interactions, indicating stable performance across these variables. Moreover, while no individual evaluation metric, such as BERTScore, BLEURT or cosine similarity, showed a high correlation with human evaluations. However, the ACC-Sim metric, which averages accuracy and cosine similarity, exhibits the highest correlation with human evaluation, with Spearman`s ρ at 0.86 and a P value < 0.001, indicating a significant ordinal correlation between ACC-Sim and human evaluation. This suggests its potential as a surrogate for human evaluation in similar medical scenarios. Conclusion: Our study did not find a significant difference between the RAG and Cold Turkey methods in terms of ChatGPT-4`s clinical decision-making performance in Afib and CIED infection management. The ACC-Sim metric closely aligns with human evaluations in these specific medical contexts and shows promise for integration into a semi-automatic evaluation framework.