Cor Vasa 2024, 66(1):37-43 | DOI: 10.33678/cor.2023.096

Artificial intelligence in resting ECG: Higher accuracy in the interpretation of rhythm abnormalities

Štěpán Havránek^a, Barbora Steklá^a, Michaela Veselá^a, Josef Holub^a, Michaela Zemková^a, Lucie Miksová^a, Karolína Kvasničková^a, Nikol Kubínová^a, Jean-Claude Lubanda^a, Milan Dusík^a, Josef Marek^a, Vladyslava Čeledová^b, Lenka Plačková^b

^a 2nd Department of Internal Cardiovascular Medicine, The First Faculty of Medicine, Charles University and General University Hospital in Prague, Prague, the Czech Republic

^b BTL Industries Ltd., Stevenage, the United Kingdom

Objective: This study aimed to evaluate the performance of a developed novel AI-based ECG rhythm model (AI-ECGRM) in binary classification between sinus rhythm and arrhythmias.

Methods: The interpretations generated by the AI-ECGRM were compared to the diagnostic conclusions made by cardiologists. The confusion matrix was used to verify the AI-ECGRM's sensitivity, specificity, positive predictive value, and negative predictive value.

Results: The testing dataset included 1,491 randomly selected ECGs (mean age 65±21 years; 54% female). Out of the testing dataset, the highly advanced cardiologists diagnosed 1,271 ECGs as sinus rhythm and 220 as arrhythmia. The AI-ECGRM labelled 1,169 as sinus rhythm and 322 as arrhythmia out of the same ECGs. The sensitivity and specificity of the model were 94% and 91%, respectively. The positive predictive value was 64%. The negative predictive value was 99%, indicating a very low probability of missing any potential pathology.

Conclusion: The results demonstrated the efficacy of the developed AI-ECGRM in accurately discriminating between ECGs exhibiting normal sinus rhythm and those indicating cardiac arrhythmias. Moreover, the AI-ECGRM exhibited an exceptional negative predictive value, approaching 100%.

Keywords: Arrhythmia, Artificial intelligence, Cardiovascular diseases, ECG interpretation, Resting ECG, Sinus rhythm

Received: December 10, 2023; Revised: December 10, 2023; Accepted: December 15, 2023; Prepublished online: June 2, 2012; Published: March 5, 2024  Show citation