Can ChatGPT accurately detect atrial fibrillation using smartwatch ECG?

IF 2.6 4区医学 Q2 CARDIAC & CARDIOVASCULAR SYSTEMS

Heart & Lung Pub Date : 2025-05-08 DOI:10.1016/j.hrtlng.2025.04.032

Yuval Avidan , Ibrahim Naoum , Razi Khoury , Sameha Zahra , Nissan Ben Dov , Jorge E Schliamser , Asaf Danon , Amir Aker

{"title":"Can ChatGPT accurately detect atrial fibrillation using smartwatch ECG?","authors":"Yuval Avidan , Ibrahim Naoum , Razi Khoury , Sameha Zahra , Nissan Ben Dov , Jorge E Schliamser , Asaf Danon , Amir Aker","doi":"10.1016/j.hrtlng.2025.04.032","DOIUrl":null,"url":null,"abstract":"<div><h3>Background</h3><div>Current guidelines require physician confirmation for smartwatch-diagnosed atrial fibrillation (AF), increasing telemedicine workloads. The newest ChatGPT-4o (GPT-4o) incorporates advanced image input capabilities.</div></div><div><h3>Objective</h3><div>To assess GPT-4o’s performance in identifying AF from smartwatch recordings.</div></div><div><h3>Methods</h3><div>Consecutive 120 patients with AF and 60 controls with sinus rhythm (SR), confirmed by conventional 12-lead ECG, recorded single-lead ECGs using an Apple Watch (AW) Series 6®. Two blinded cardiologists independently classified the smartwatch recordings as AF, SR, or undetermined. GPT-4o was subsequently prompted to analyze all smartwatch ECGs.</div></div><div><h3>Results</h3><div>Six AF cases were excluded due to undetermined AW-ECG recordings, leaving 114 AF patients (mean age: 73.4 ± 10.4 years) and 60 controls. The AW algorithm achieved 97.3 % and 100 % accuracy for AF and SR, respectively, while GPT-4o correctly analyzed 47.3 % of AF and 71.6 % of SR tracings. None of the AF characteristics—chronicity, heart rate, QRS width, fibrillatory wave amplitude, or R-wave amplitude and polarity—were predictive of GPT-4o's diagnostic accuracy.</div></div><div><h3>Conclusion</h3><div>The current capabilities of GPT-4o are insufficient to make a reliable diagnosis of AF from smartwatch ECGs. Despite the theoretical appeal of leveraging this innovative technology for such purpose, the findings highlight that human expertise remains indispensable. Consumers must remain aware of the current limitations of this technology.</div></div>","PeriodicalId":55064,"journal":{"name":"Heart & Lung","volume":"73 ","pages":"Pages 90-94"},"PeriodicalIF":2.6000,"publicationDate":"2025-05-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Heart & Lung","FirstCategoryId":"3","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0147956325001062","RegionNum":4,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"CARDIAC & CARDIOVASCULAR SYSTEMS","Score":null,"Total":0}

引用次数: 0

Abstract

Background

Current guidelines require physician confirmation for smartwatch-diagnosed atrial fibrillation (AF), increasing telemedicine workloads. The newest ChatGPT-4o (GPT-4o) incorporates advanced image input capabilities.

Objective

To assess GPT-4o’s performance in identifying AF from smartwatch recordings.

Methods

Consecutive 120 patients with AF and 60 controls with sinus rhythm (SR), confirmed by conventional 12-lead ECG, recorded single-lead ECGs using an Apple Watch (AW) Series 6®. Two blinded cardiologists independently classified the smartwatch recordings as AF, SR, or undetermined. GPT-4o was subsequently prompted to analyze all smartwatch ECGs.

Results

Six AF cases were excluded due to undetermined AW-ECG recordings, leaving 114 AF patients (mean age: 73.4 ± 10.4 years) and 60 controls. The AW algorithm achieved 97.3 % and 100 % accuracy for AF and SR, respectively, while GPT-4o correctly analyzed 47.3 % of AF and 71.6 % of SR tracings. None of the AF characteristics—chronicity, heart rate, QRS width, fibrillatory wave amplitude, or R-wave amplitude and polarity—were predictive of GPT-4o's diagnostic accuracy.

Conclusion

The current capabilities of GPT-4o are insufficient to make a reliable diagnosis of AF from smartwatch ECGs. Despite the theoretical appeal of leveraging this innovative technology for such purpose, the findings highlight that human expertise remains indispensable. Consumers must remain aware of the current limitations of this technology.

Abstract Image

查看原文本刊更多论文

ChatGPT能否通过智能手表心电准确检测心房颤动？

目前的指南要求医生确认智能手表诊断的心房颤动（AF），增加了远程医疗工作量。最新的chatgpt - 40 （gpt - 40）集成了先进的图像输入功能。目的评估gpt - 40在智能手表记录识别AF中的性能。方法连续120例房颤患者和60例有窦性心律（SR）的对照组，经常规12导联心电图确认，使用Apple Watch (AW) Series 6®记录单导联心电图。两名盲法心脏病专家独立地将智能手表记录归类为AF、SR或不确定。随后，gpt - 40被提示分析所有智能手表的心电图。结果6例房颤患者因AW-ECG记录不确定而被排除，剩下114例房颤患者（平均年龄：73.4±10.4岁）和60例对照组。AW算法对AF和SR的准确率分别达到97.3%和100%，而gpt - 40对AF和SR的准确率分别为47.3%和71.6%。房颤特征——慢性、心率、QRS宽度、纤颤波振幅或r波振幅和极性——均不能预测gpt - 40的诊断准确性。结论gpt - 40目前的能力不足以通过智能手表心电图对房颤进行可靠的诊断。尽管利用这种创新技术实现这一目的在理论上很有吸引力，但研究结果强调，人类的专业知识仍然是不可或缺的。消费者必须意识到这项技术目前的局限性。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

Heart & Lung 医学-呼吸系统

CiteScore

4.60

自引率

3.60%

发文量

184

审稿时长

35 days

期刊介绍： Heart & Lung: The Journal of Cardiopulmonary and Acute Care, the official publication of The American Association of Heart Failure Nurses, presents original, peer-reviewed articles on techniques, advances, investigations, and observations related to the care of patients with acute and critical illness and patients with chronic cardiac or pulmonary disorders. The Journal''s acute care articles focus on the care of hospitalized patients, including those in the critical and acute care settings. Because most patients who are hospitalized in acute and critical care settings have chronic conditions, we are also interested in the chronically critically ill, the care of patients with chronic cardiopulmonary disorders, their rehabilitation, and disease prevention. The Journal''s heart failure articles focus on all aspects of the care of patients with this condition. Manuscripts that are relevant to populations across the human lifespan are welcome.