Heart Lung. 2025 Sep 13:S0147-9563(25)00182-7. doi: 10.1016/j.hrtlng.2025.08.007. Online ahead of print.
ABSTRACT
BACKGROUND: DeepSeek is a recently launched large language model (LLM), whereas GPT-4o is an advanced ChatGPT version whose electrocardiography (ECG) interpretation capabilities have been previously studied. However, DeepSeek’s performance in this domain remains unexplored.
OBJECTIVES: This study aims to evaluate DeepSeek’s accuracy in ECG interpretation and compare it with GPT-4o, emergency medicine specialists, and cardiologists. A secondary aim is to assess any performance changes in GPT-4o over one year.
METHODS: Between February 9 and March 1, 2025, 40 ECG images (20 daily routine, 20 more challenging) from the book 150 ECG Cases were evaluated by both GPT-4o and DeepSeek, each model tested 13 times. The accuracy of their responses was compared with previously collected answers from 12 cardiologists and 12 emergency medicine specialists. GPT-4o’s 2025 performance was compared to its 2024 results on identical ECGs.
RESULTS: GPT-4o outperformed DeepSeek with higher median correct answers on daily routine (14 vs. 12), more challenging (13 vs. 10), and total ECGs (27 vs. 22) with statistically significant differences (p=0.048, p<0.001, p<0.001). A moderate agreement was observed between the responses provided by GPT-4o (p<0.001, Fleiss Kappa=0.473), while a substantial agreement was observed in the responses provided by DeepSeek (p<0.001, Fleiss Kappa=0.712). No significant year-over-year improvement was observed in GPT-4o’s performance.
CONCLUSION: This first evaluation of DeepSeek in ECG interpretation reveals its performance is lower than that of GPT-4o and human experts. While GPT-4o demonstrates greater accuracy, both models fall short of expert-level performance, underscoring the need for caution and further validation before clinical integration.
PMID:40947358 | DOI:10.1016/j.hrtlng.2025.08.007
