Clin Res Cardiol (2022). https://doi.org/10.1007/s00392-022-02002-5
Prediction of Cardiovascular Risk Factors and Cardiovascular Diseases in a Population-Based Cohort Study from the ECG using Deep Learning – Pilot Data from an AI Imaging Pipeline
M. Knorr1, M. Neyazi2, J. P. Bremer1, M. Vollmer3, B. Schrage2, K. Surendra1, J. Brederecke1, F. M. Ojeda1, S. Nürnberg4, S. Groß5, M. Dörr5, S. Blankenberg6, R. Schnabel7
1Universitäres Herz- und Gefäßzentrum Hamburg GmbH, Hamburg; 2Klinik für Kardiologie, Universitäres Herz- und Gefäßzentrum Hamburg GmbH, Hamburg; 3Institut für Bioinformatik, Universität Greifswald, Greifswald; 4Universitätsklinikum Hamburg-Eppendorf Institut für Angewandte Medizininformatik, Hamburg; 5Klinik und Poliklinik für Innere Medizin B, Universitätsmedizin Greifswald, Greifswald; 6Klinik für Kardiologie, Universitäres Herz- und Gefäßzentrum UKE Hamburg GmbH, Hamburg; 7Allgemeine und Interventionelle Kardiologie, Universitäres Herz- und Gefäßzentrum Hamburg GmbH, Hamburg;

Introduction: An earlier identification of individuals who are at risk of developing cardiovascular diseases enables a more timely initiation of medical intervention and prophylaxis. We sought to build deep learning-based screening algorithms based on the electrocardiogram (ECG) employing data from a large population-based cohort study with a focus on CHA2DS2-VASc score risk factors.

Methods:  We utilized data from the first 8489 participants (median age 62.0 years, 49.6% men) with 12-channel-ECGs from the population-based Hamburg City Health Study (HCHS) and corresponding risk factors diabetes, heart failure, hypertension, coronary heart disease and history of stroke. We applied a 1D-Convolutional-Transformer deep learning model for predicting these risk factors from the ECGs. First, we split the dataset in a large hold-out test set (n = 2932) to determine the final  area under the receiver operating curve scores (AUC). The remaining (n=5557) are deployed in a 5-fold cross validation scheme. We compare our results to a 5-fold conventional statistic baseline linear regression using age and sex as features, which was fitted in the same manner on the same data.

Results: Our model shows good performance predicting diabetes (AUC 0.67 confidence interval (CI) 0.66-0.69), heart failure (AUC 0.73, CI 0.71-0.76), hypertension (AUC 0.69, CI 0.68-0.71), and coronary artery disease (AUC 0.67, CI 0.66-0.68) and moderate results on predicting a history of stroke (AUC 0.61, CI 0.57-0.64). The ranked average of the 5-folds yields even better results for diabetes (AUC 0.70), heart failure (AUC 0.75), hypertension (AUC 0.71), coronary artery disease (AUC 0.69) and history of stroke (AUC 0.62) (Fig. 1). This shows better performance than the baseline linear regression model for heart failure (AUC 0.64), hypertension (AUC 0.70) and diabetes (AUC 0.66), but not for coronary artery disease (AUC 0.72) and history of stroke (AUC 0.67) (Fig. 1).

Conclusions: Prevalent cardiovascular disease and cardiac risk factors can be predicted from the ECG to a certain extent. Future research will be focused on whether these predicted risk factors are relevant in the prediction of incident cardiac events as they may reflect early cardiac damage.
https://dgk.org/kongress_programme/jt2022/aV981.html