Artificial intelligence in resting ecg: higher accuracy in the interpretation of rhythm abnormalities

Abstract Introduction The electrocardiogram (ECG) examination is part of a daily routine in primary care, and non-cardiologist physicians deal with high misinterpretations of ECGs. To improve and expand the idea of automated ECG interpretation and support non-cardiologist physicians in ECG interpretation, a new AI-based ECG rhythm model (AI-ECGRM) was developed. The study aimed to test the AI-ECGRM for binary classification between sinus rhythm (SR) and arrhythmias (ARs). Methods The confusion matrix was used to verify the AI-ECGRM’s sensitivity, specificity, and positive and negative predictive values. Two physicians (one cardiologist) evaluated all ECGs used for testing. Results The testing dataset included 1,491 randomly selected ECGs (mean age 65±21 years; 54% female). Out of the testing dataset, the cardiologists diagnosed 1,271 ECGs as SR and 220 as AR. The AI-ECGRM labeled 1,169 as SR and 322 as AR out of the same ECGs. The model's sensitivity and specificity were 94% and 91%, respectively. The positive predictive value was 64%, and the negative predictive value was 99%. The high proportion of false positive results was caused mainly by ECGs with abnormalities other than ARs or by technical artifacts. The false negativity was given by missing a single extrasystoles by model. Conclusion The developed AI-ECGRM proved to be effective in distinguishing between the ECGs with regular SR and those with cardiac ARs, while the negative predictive value of the AI-ECGRM was nearly 100%.