Prognostic Significance and Associations of Neural Network-Derived Electrocardiographic Features

Abstract

Background: Subtle, prognostically important ECG features may not be apparent to physicians. In the course of supervised machine learning, thousands of ECG features are identified. These are not limited to conventional ECG parameters and morphology. We aimed to investigate whether neural network-derived ECG features could be used to predict future cardiovascular disease and mortality and have phenotypic and genotypic associations.

Methods: We extracted 5120 neural network-derived ECG features from an artificial intelligence-enabled ECG model trained for 6 simple diagnoses and applied unsupervised machine learning to identify 3 phenogroups. Using the identified phenogroups, we externally validated our findings in 5 diverse cohorts from the United States, Brazil, and the United Kingdom. Data were collected between 2000 and 2023.

Results: In total, 1 808 584 patients were included in this study. In the derivation cohort, the 3 phenogroups had significantly different mortality profiles. After adjusting for known covariates, phenogroup B had a 20% increase in long-term mortality compared with phenogroup A (hazard ratio, 1.20 [95% CI, 1.17-1.23]; P<0.0001; phenogroup A mortality, 2.2%; phenogroup B mortality, 6.1%). In univariate analyses, we found phenogroup B had a significantly greater risk of mortality in all cohorts (log-rank P<0.01 in all 5 cohorts). Phenome-wide association study showed phenogroup B had a higher rate of future atrial fibrillation (odds ratio, 2.89; P<0.00001), ventricular tachycardia (odds ratio, 2.00; P<0.00001), ischemic heart disease (odds ratio, 1.44; P<0.00001), and cardiomyopathy (odds ratio, 2.04; P<0.00001). A single-trait genome-wide association study yielded 4 loci. SCN10A, SCN5A, and CAV1 have roles in cardiac conduction and arrhythmia. ARHGAP24 does not have a clear cardiac role and may be a novel target.

Conclusions: Neural network-derived ECG features can be used to predict all-cause mortality and future cardiovascular diseases. We have identified biologically plausible and novel phenotypic and genotypic associations that describe mechanisms for the increased risk identified.

Keywords: cardiovascular diseases; electrocardiography; neural networks, computer; supervised machine learning; unsupervised machine learning.

Figure 1.

Data analysis pipeline. A, Hybrid machine learning approach with a combination of supervised and unsupervised machine learning to use neural network–derived ECG features to identify phenogroups from the 12-lead ECG. Three-dimensional figure shows a dimensionally reduced representation of the phenogroups. B, Flow of data and analyses performed in this study. AF indicates atrial fibrillation; AVB, atrioventricular block; BIDMC, Beth Israel Deaconess Medical Center; BN, batch normalization; Conv, convolutional layer; CODE, Clinical Outcomes in Digital Electrocardiography; ELSA-Brasil, Brazilian Longitudinal Study of Adult Health; Grad-CAM, gradient-weighted class activation mapping; GWAS, genome-wide association study; LBBB, left bundle branch block; ML, machine learning; PheWAS, phenome-wide association study; ReLu, rectified linear unit; ResBlk, residual block; RBBB, right bundle branch block; and SaMi-Trop, São Paulo-Minas Gerais Tropical Medicine Research Center.

Figure 2.

Survival analysis in the derivation data set. The 3 phenogroups have prognostic significance, with phenogroup B having a markedly worse prognosis. Data are shown for (A) the whole Clinical Outcomes in Digital Electrocardiography (CODE) cohort and (B) the CODE cohort with the removal of subjects with any of the following diagnoses on the ECG: first-degree atrioventricular block, right bundle branch block, left bundle branch block, sinus tachycardia, sinus bradycardia, and atrial fibrillation. C, Subset of the CODE cohort with normal ECGs. CODE-CNN indicates CODE-convolutional neural network.

Figure 3.

Survival analysis in the 5 external validation data sets. In the volunteer populations (Whitehall II, UK Biobank, and Brazilian Longitudinal Study of Adult Health (ELSA-Brasil), phenogroup C has few subjects and, therefore, is excluded. Phenogroup B has a significantly higher event rate. A, Whitehall II cohort and (B) UK Biobank, survival free of major adverse cardiovascular events is depicted. C, ELSA-Brasil cohort; D, São Paulo-Minas Gerais Tropical Medicine Research Center (SaMi-TROP) cohort; and E, Beth Israel Deaconess Medical Center (BIDMC) cohort.

Figure 4.

Phenome-wide association study (PheWAS). Manhattan plot showing negative logarithm of the univariate correlation P value between phenotypes for Beth Israel Deaconess Medical Center (BIDMC) disease phecodes, Disease PheWAS (A), and correlation coefficient for UK Biobank phenotypes, Biobank PheWAS (B). Small points depict associations not reaching statistical significance, while large points crossed the Bonferroni threshold for statistical significance. BMC indicates bone mineral content; IMT, intima-media thickness; LAV, left atrial volume; LV, left ventricle; LVEDV, left ventricular end-diastolic volume; LVESV, left ventricular end-systolic volume; MRI, magnetic resonance imaging; RAV, right atrial volume; RV, right ventricle; and RVESV, left ventricular end-systolic volume.

Figure 5.

Genome-wide association study. Manhattan plots of genomic loci associated with ECG phenogroup. The nearest genes are annotated on the plot. The red line depicts the genome-wide significant threshold (P<5×10⁻⁸). AF indicates atrial fibrillation; DCM, dilated cardiomyopathy; and SCD, sudden cardiac death.

Figure 6.

Model explainability. Gradient-weighted class activation mapping is used to generate importance maps showing the sections of the ECG signal deemed most important for phenogroup determination. The average saliency of 1000 ECGs from the center of each cluster is shown. Areas marked with green show the terminal QRS and terminal T wave are important for identification of the high-risk phenogroup (phenogroup B).