Abstract Sat305: Race and Ethnicity Classification via Deep Learning Electrocardiogram Analysis

Background: Cardiovascular diseases are a leading cause of morbidity and mortality worldwide. Demographics, including race/ethnicity, are among the strongest risk factors for predicting cardiovascular events, including cardiac arrest. Thus, the electrocardiogram (ECG) could serve as a surrogate for risk assessment, if the ECG can reliably classify according to race/ethnicity. This is particularly important when demographic information is not available, such as with datasets accumulated from wearable devices. Hypothesis: It was hypothesized that ECG data, in conjunction with deep learning algorithms, could be used to classify individuals by race/ethnicity. Methods: 10 seconds of 12-lead ECG voltage waveform data from the Nightingale Open Science - Assessing Heart Attack Risk dataset were utilized to develop a convolutional neural network model capable of classifying individuals by race and ethnicity. The cohort included 6,168 black, 4,885 Hispanic, and 16,982 white individuals. Subsets of the data were used to train, validate, and test three models. The first model classified black as compared to the other two groups (B-HW), the second Hispanic compared to the others (H-BW), and the third White compared to the others (W-BH). Model training was performed using 9-fold cross-validation with ten epochs each. Model performance was evaluated via area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, and specificity. One-way ANOVA and post-hoc Dunn’s tests were used to determine differences and groupings. Results: Using AUROC, black and Hispanic individuals were classified more correctly than white individuals, and these differences were significant (p<0.001) (Figure 1). More specifically, average AUROC, accuracy, sensitivity and specificity for black individuals were 0.823, 78.1%, 78.2%, and 99.7%, respectively, while for Hispanic individuals, values were 0.839, 82.3%, 82.4%, and 99.9%, respectively, and for white individuals, metrics were 0.677, 63.3%, 63.2%, and 18.0%, respectively. Conclusion: ECGs analyzed with deep learning independently explained race and ethnicity with reasonable reliability. ECGs from black and Hispanic individuals are more reliably predicted than those from white individuals. Thus, this study supports ECGs and deep learning as potentially valuable and complementary tools for determining cardiovascular risk when race/ethnicity data are not available.