Clin Res Cardiol (2022). https://doi.org/10.1007/s00392-022-02002-5

Predicting NT-proBNP as a marker for Myocardial Stress from the ECG using deep learning – Pilot Data from an AI Imaging Pipeline

M. Neyazi1, J. P. Bremer1, M. Knorr1, M. Vollmer2, B. Schrage1, K. Surendra1, J. Brederecke1, F. M. Ojeda1, S. Nürnberg3, S. Groß4, M. Dörr4, S. Blankenberg1, R. Schnabel1

1Klinik für Kardiologie, Universitäres Herz- und Gefäßzentrum UKE Hamburg GmbH, Hamburg; 2Institute of Bioinformatics, Greifswald; 3Institut für Angewandte Medizininformatik, Hamburg; 4Klinik und Poliklinik für Innere Medizin B, Universitätsmedizin Greifswald, Greifswald;

Introduction: The cardiac peptide hormone brain natriuretic peptide (NT-proBNP) is physiologically released in the left ventricle and atria upon myocardial stretch. As alterations in the heart affect its electrophysiology, we expect that even minimal changes in the electrocardiogram (ECG) can be indicative for the same underlying conditions which result in elevated NT-proBNP concentrations. Utilizing machine learning models and a large sample of a population-based cohort, we examined the relation between digital ECG information and NT-proBNP for an improved understanding of heart failure and a potential screening tool for the general population. Methods: Data from the Hamburg City Health Study, which is a population-based cohort study, was used. The first data release provides 12-channel-ECGs of 8256 participants (median age 62.0 years, 49.6% men) and corresponding NT-proBNP concentrations (mean±standard deviation 144±385 pg/ml). Here we employ a 1D-Convolutional-Transformer deep learning model for predicting NT-proBNP levels from raw ECGs of 10 seconds length. First, we split the dataset into a hold-out test dataset (n = 2860), which all metrics are calculated on. The rest of the dataset (n = 5416) is employed in a 5-fold validation scheme. We compare our results to a 5-fold linear regression baseline using age and sex as features. Lastly, we compare the area under the receiver operating curve (AUC) for the diagnosis of heart failure from the predicted NT-proBNP levels to a model trained to classify heart failure directly. Results: Our model shows good performance predicting NT-proBNP values on the 5-fold validation scheme (R-squared [R²] 0.26, confidence interval [CI] 0.24-0.28, mean absolute error [mae] 83, CI 82-84) and the average prediction of all 5-folds yields even better performance (R² 0.27, mae 81) (Fig. 1).  Our baseline linear regression model in comparison shows significantly worse performance on the 5-fold validation scheme (R² 0.07, CI 0.05-0.09, mae 111, CI 102-120) and the mean prediction of all 5-folds (R² 0.07, mae 111). Next, we have used the ranked average of our predicted NT-proBNP levels of our 5 models from the k-fold validation scheme as a predictor for heart failure. Our predicted NT-proBNP levels show high performance for the classification of heart failure (AUC 0.82), while the model classifying heart failure directly has a lower AUC of 0.75. (Fig. 2). Conclusions: Values of NT-proBNP can, to a certain extent, be directly inferred using an ECG sample and hold potential to be used for screening of the general population for heart failure. In comparison to a simple baseline model on the participants’ heart failure diagnosis we show much improved performance, leading to the conclusion that the ECG reveals changes which correlate with elevated NT-proBNP level and chronic heart failure.

https://dgk.org/kongress_programme/jt2022/aV11.html