Search papers, labs, and topics across Lattice.
This paper investigates the reliability of GPT-4o in applying the EHRA AI in EP checklist to assess AI research in cardiac electrophysiology. The authors compared GPT-4o's assessments of 55 manuscripts against those of an expert panel, revealing significant inconsistencies. The study concludes that while LLMs can assist in manuscript evaluation, their current limitations necessitate human oversight to ensure compliance with reporting standards.
GPT-4o can't reliably review AI-in-cardiology papers yet, disagreeing with expert reviewers on 7 out of 29 checklist items per paper.
Abstract Background The growing potential of Artificial Intelligence (AI) in cardiac electrophysiology (EP) led to an increasing number of publications exploring possible areas of AI-application in clinical settings. To improve reader understanding and to standardize reporting of AI-related research, the EHRA AI checklist was recently published. For its validation, studies from three EP domains were applied to the checklist by an expert panel. Aim of this study was to explore whether AI could assist in analysing and evaluating scientific manuscripts. Methods The study was conducted in May 2025. We used the Large Language Model (LLM) ChatGPT (GPT-4o) by OpenAI to apply all studies from the EHRA AI manuscript to the checklist and compared the generated responses by the LLM with those provided by the expert panel. Firstly, after ensuring that each item from the checklist (PDF file) could accurately be extracted by ChatGPT, we asked the LLM to create its own prompt which was then further refined by us through prompt engineering. Secondly, each study from the EHRA AI manuscript together with the checklist was successively uploaded to ChatGPT. For each analysis, the prompt remained the same: (1) asking the LLM to act as a reviewer in clinical EP, (2) short explanation of the two uploaded files, (3) a description of the task to decide for each item of the checklist whether it is "applicable" or "not applicable", (4) a description of the desired output format and (4) additional rules such as to not change or remove any item from the list. Results All 29 items from the checklist were analysed using one prompt for each of the 55 manuscripts. Results of the initial LLM assessment compared to the initial expert assessment are shown in the Figure. Per manuscript, none of ChatGPT’s results matched the human results completely, and the mean number of inconsistencies between human and LLM rating was 7 ± 2.8. Per checklist item, only item 11, which refers to the description of the data type used in the study, was rated identically to the human reviewer for every manuscript. Per EHRA AI checklist item, the mean number of inconsistencies among all papers was 13 ± 9.0. Conclusion Standards for reporting and understanding AI in EP are important. Even if LLMs can help with the evaluation and analysis of scientific manuscripts, there are currently still limitations in terms of reliability, meaning that human intelligence is still required to comply with standards. Comparison of human and LLM ratings
