Deep learning-based NT-proBNP prediction from the ECG for risk assessment in the community.

OBJECTIVES: The biomarker N-terminal pro B-type natriuretic peptide (NT-proBNP) has predictive value for identifying individuals at risk for cardiovascular disease (CVD). However, it is not widely used for screening in the general population, potentially due to financial and operational reasons. This study aims to develop a deep-learning model as an efficient means to reliably identify individuals at risk for CVD by predicting serum levels of NT-proBNP from the ECG. METHODS: A deep convolutional neural network was developed using the population-based cohort study Hamburg City Health Study (HCHS, n=8,253, 50.9 % women). External validation was performed in two independent population-based cohorts (SHIP-START, n=3,002, 52.1 % women, and SHIP-TREND, n=3,819, 51.2 % women). Assessment of model performance was conducted using Pearson correlation (R) and area under the receiver operating characteristics curve (AUROC). RESULTS: NT-proBNP was predictable from the ECG (R, 0.566 [HCHS], 0.642 [SHIP-START-0], 0.655 [SHIP-TREND-0]). Across cohorts, predicted NT-proBNP (pNT-proBNP) showed good discriminatory ability for prevalent and incident heart failure (HF) (baseline: AUROC 0.795 [HCHS], 0.816 [SHIP-START-0], 0.783 [SHIP-TREND-0]; first follow-up: 0.669 [SHIP-START-1, 5 years], 0.689 [SHIP-TREND-1, 7.3 years]), comparable to the discriminatory value of measured NT-proBNP. pNT-proBNP also demonstrated comparable results for other incident CVD, including atrial fibrillation, stroke, myocardial infarction, and cardiovascular death. CONCLUSIONS: Deep learning ECG algorithms can predict NT-proBNP concentrations with high diagnostic and predictive value for HF and other major CVD and may be used in the community to identify individuals at risk. Long-standing experience with NT-proBNP can increase acceptance of such deep learning models in clinical practice.

Pragmatic screening for heart failure in the general population using an electrocardiogram-based neural network.

AIMS: We aim to develop a pragmatic screening tool for heart failure at the general population level. METHODS AND RESULTS: This study was conducted within the Hamburg-City-Health-Study, an ongoing, prospective, observational study enrolling randomly selected inhabitants of the city of Hamburg aged 45-75 years. Heart failure was diagnosed per current guidelines. Using only digital electrocardiograms (ECGs), a convolutional neural network (CNN) was built to discriminate participants with and without heart failure. As comparisons, known risk variables for heart failure were fitted into a logistic regression model and a random forest classifier. Of the 5299 individuals included into this study, 318 individuals (6.0%) had heart failure. Using only the digital ECGs instead of several risk variables as an input, the CNN provided a comparable predictive accuracy for heart failure versus the logistic regression model and the random forest classifier [area under the curve (AUC) of 0.75, a sensitivity of 0.67 and a specificity of 0.69 for the CNN; AUC 0.77, a sensitivity of 0.63 and a specificity of 0.76 for the logistic regression; AUC 0.79, a sensitivity of 0.67 and a specificity of 0.72 for the random forest classifier]. CONCLUSIONS: Using a CNN build on digital ECGs only and requiring no additional input, we derived a screening tool for heart failure in the general population. This could be perfectly embedded into clinical routine of general practitioners, as it builds on an already established diagnostic tool and does not require additional, time-consuming input. This could help to alleviate the underdiagnosis of heart failure.