Misclassification of sex by deep neural networks reveals novel ECG characteristics that explain a higher risk of mortality in women and in men. (14th October 2021)

• Record Type: Journal Article
• Authors: Siegersma, K; Van De Leur, R; Onland-Moret, N C; Van Es, R; Den Ruijter, H M
• Is Part Of: European heart journal. Volume 42(2021)Supplement 1
• Publication Date: 2021-10-14
• Subjects: Big Data Analysis; Cardiology -- Periodicals; Heart -- Diseases -- Periodicals; 616.12005
• DOI: 10.1093/eurheartj/ehab724.3162
• Languages: English
• ISSNs: 0195-668X

Abstract

Abstract: Background: Performing sex-stratified analyses in medical research can lead to new insights. Artificial intelligence is increasingly used on electrocardiograms (ECGs) for prediction of mortality, risk and diagnosis. ECG-based deep neural networks (DNN) have shown to be able to distinguish women from men. This classification inevitably leads to a misclassified group. It is unknown what ECG characteristics account for sex classification, and how these variables affect mortality. We hypothesize that misclassification of sex by a DNN on ECGs can lead to new insights on ECG variables and mortality in women and men. Aim: To study if DNN-based sex classification and misclassification identifies new ECG features associated with mortality in women and in men. Methods: All ECGs spanning three decades from our University Medical Center were selected (n=1.136.113). A DNN was trained to classify sex based on 12-lead ECG using 131.673 normal ECGs of 68.500 subjects (48.6% women). Validation was performed on the other half of the population (68.500 ECGs, 49.5% women). Correctly classified and misclassified women and men were grouped. Discriminatory performance of the DNN was assessed using the AUC. The DNN was used to assess which characteristics influenced classification. We post-hoc tested their association with misclassification. To assess the association between sex-classification and mortality, time-to-event analysis was done with Kaplan-Meier curves. All individuals from 18 to 85 years, with at least 1 year follow-up were selected. To assess which ECG characteristics explain differences in mortality between the groups, sex-specific mediation analysis was performed using Weibull regression. Results: DNN classification could distinguish women and men based on ECG (AUC: 0.97, 95% CI: 0.9789–0.9806). However, still 2.589 (8.1%) men and 2.368 (7.5%) women were misclassified. During a median follow-up of 8 years and 10 months, 4066 (13.0%) men died and 3055 (10.2%) women died, of whom, respectively 380 (9.3%) and 318 (9.9%) were misclassified. Misclassified individuals had worse survival than their correctly classified biological peers (misclassified vs correct classified women HR: 1.31, 95% CI: 1.17–1.48, misclassified vs correct classified men HR: 1.36, 95% CI: 1.22–1.51, figure 1). Mediation analysis showed that in men known ECG variables could partly explain the association between misclassification of sex and mortality. In women this was less so (figure 2). Moreover, this study revealed a new causal relation between QRS-shortening and mortality in women (figure 2). Conclusion: A DNN can accurately classify women and men based on ECGs. Misclassification of sex is associated with worse survival in both sexes, and explained by sex-specific ECG features such as QRS shortening in women. This novel finding underscores the importance of integrating sex in AI to uncover previously unknown associations with mortality, and to prevent bias in algorithms. Funding Acknowledgement: Type of funding sources: Public grant(s) – National budget only. Main funding source(s): CVON-AIZonMW