Machine learning methods, particularly a convolutional neural network, accurately classified 12-lead ECGs in young people according to the International Criteria during real-time acquisition.
Cross-Sectional (n=22,488)
Do machine learning methods accurately classify 12-lead ECGs in young people according to the International Criteria compared to cardiologist interpretation?
Machine learning models, particularly convolutional neural networks, show promise for automated, accurate classification of screening ECGs in young individuals.
Abstract Introduction Artificial intelligence-enhanced electrocardiogram (AI-ECG) models represent promising tools for enhancing the ECG as a diagnostic and prognostic tool for cardiovascular disease, including conditions associated with sudden cardiac death (SCD) in young individuals. Whilst an emerging role for the implementation of AI-ECG in sports cardiology has been described, the body of evidence for its use in cardiac screening is limited. Purpose To determine whether machine learning (ML) methods can accurately and automatically classify 12-lead ECGs in young people according to the International Criteria for ECG interpretation in athletes. Methods A total of 22,488 ostensibly healthy individuals aged 14-35 years (41% female) who underwent screening with a 12-lead ECG were included. Each 10s, 500Hz 12-lead ECG was labelled by the responsible cardiologist according to the International Criteria (normal, 96%; borderline, 2%; abnormal, 2%). A random 20% (n=4,500) sample was held out for testing. Three approaches were evaluated: (i) kernel support vector machine (kSVM), (ii) XGBoost, both trained on ECG features extracted algorithmically, and (iii) a 1D Resnet-SE (10) convolutional neural network (CNN) trained directly on the ECG waveforms. Performance was evaluated using accuracy, F1-score and precision at fixed sensitivities (0.99-0.999). Results The three ML models were able to accurately characterise ECGs, as shown in Table 1. The CNN demonstrated the greatest ability to detect subtle ECG features, albeit with reduced interpretability relative to the other methods. All models operated in real time during ECG acquisition, and their performance, particularly that of the CNN, is expected to improve with additional training data. At very high sensitivities, model precision approximated the expected prevalence of clinically significant borderline (orange) and abnormal (red) findings. Conclusion Machine learning methods show strong promise for automated ECG classification in young people, consistent with their successful application in older at-risk populations. Model performance should further improve with larger training datasets and additional clinical features. With continued optimisation and external validation, AI-ECG approaches are well positioned to serve as an effective triage tool in cardiac screening, reducing inequalities by improving access at settings with limited expertise and resources at lower cost.For image description, please refer to the figure legend and surrounding text. For image description, please refer to the figure legend and surrounding text.
Tardo et al. (Mon,) conducted a cross-sectional in Cardiac screening (n=22,488). Machine learning methods (kSVM, XGBoost, CNN) vs. Cardiologist interpretation was evaluated on Accuracy, F1-score and precision at fixed sensitivities (0.99-0.999). Machine learning methods, particularly a convolutional neural network, accurately classified 12-lead ECGs in young people according to the International Criteria during real-time acquisition.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: