Prediction and classification of aortic stenosis severity based on electrocardiogram with deep learning model

Abstract Background The prognosis of aortic stenosis (AS) is significantly affected by severity, and it is determined by echocardiography. Therefore, AS is often misdiagnosed due to not easy access to echocardiography, which requiring expense and long-term training. Recent advancements have confirmed the feasibility of diagnosing AS through electrocardiogram (ECG) analysis utilizing artificial intelligence (AI). Nonetheless, the prediction of AS severity remains a relevant challenge. The present study aimed to develop an AI model capable of predicting AS severity based on ECG, facilitating early adaptation for interventional therapy. Methods Echocardiographic data was collected from 158,399 patients between Jan 2013 to Dec 2022. 12-lead ECGs from 85,975 patients were matched with echocardiography. AS severity was determined by echocardiography findings. A total of 54,738 patients were excluded: deviation of time interval more than 180-day between ECG and echocardiogram, missing demographic information, noisy ECG, and insufficient indexing AS severity. The AI model used for the prediction of severity consisted of a convolutional neural network with bottleneck structure as a backbone and explainable boosting machine as a head model. We divided the dataset into training, validation, and test sets as 6:2:2 ratio. Results Model performance for each AS severity was 84.5% of mMAPE (maximum Mean Absolute Percentage Error). For the clinical application, AS severity was classified into binary group by thresholding at each severity. Prediction model achieved the area under the receiver operating characteristic curve (AUC) as 0.823 (95% CI: 0.805-0.841), 0.901 (95% CI: 0.867-0.934), 0.909 (95% CI: 0.881-0.937), 0.926 (95% CI: 0.908-0.945), 0.936 (95% CI: 0.913-0.959) when the threshold is 0.5 (aortic sclerosis), 1.0 (mild AS), 2 (mild to moderate AS), 3 (moderate AS), 3.5 (moderate to severe AS), respectively. When using a threshold of 3.5 (moderate to severe AS) to identify patients requiring surgical or interventional treatment, model performance achieved AUC as 0.936 (95% CI: 0.913-0.959). Conclusions The findings of this study present a promising avenue for early diagnosis and prediction for AS severity using ECGs. As threshold increases, predictive accuracy is also increased. In this prediction model, threshold can be adjusted for each medical system based on the desired sensitivity. AI prediction model with ECG might help to rapidly identify patients requiring surgical or interventional treatment.