Abstract
Cardiovascular disease (CVD) risk stratification traditionally treats sex as a binary variable, potentially masking significant biological heterogeneity. This study introduces an AI-ECG model that identifies a "sex-related cardiovascular risk continuum" through a sex discordance score. By analyzing over 1.3 million ECGs from two large cohorts, the researchers found that females with "male-like" ECG patterns face significantly higher risks of cardiovascular death and disease, even when their heart tests appear clinically normal. This AI-derived biomarker offers a path toward more personalized risk assessment and closing the sex-based gap in cardiac care.
Introduction
Cardiovascular disease remains a leading cause of death for women, yet they are frequently underserved by current medical paradigms. Traditionally, sex is treated as a binary variable (male or female) in risk stratification. This approach fails to account for biological heterogeneity within each sex. Furthermore, women are more likely to be misdiagnosed or receive less aggressive preventive care, partly because clinicians may perceive them as being at lower risk. This study seeks to move beyond the binary and determine if an AI could detect a biological "sex continuum" in electrocardiograms (ECGs) that identifies hidden cardiovascular vulnerabilities.
Methodology
The researchers developed an AI-ECG model using a Convolutional Neural Network (CNN) with residual blocks to classify sex based on 12-lead ECG data.
The Sex Discordance Score
The core innovation is the sex discordance score. While the model outputs a continuous probability (0 to 1) for sex, the discordance score is defined as the absolute difference between that prediction and the patient's self-reported biological sex. A high score indicates an ECG phenotype that deviates significantly from the average for that sex.
Explainability and Statistical Rigor
To understand the morphological features driving the AI's predictions (e.g., QRS duration or T-wave shape), the team utilized Variational Autoencoders (VAEs) and latent traversal. Cox proportional hazards models were then used to link the discordance score to long-term outcomes like mortality, heart failure, and myocardial infarction, adjusting for standard risk factors like age, smoking, and diabetes.
Datasets
The study utilized two massive, diverse cohorts to ensure robustness:
- BIDMC (Derivation/Internal Validation): Over 1.1 million ECGs from 189,539 patients at Beth Israel Deaconess Medical Center. This represented a "real-world" clinical population with various comorbidities.
- UK Biobank (External Validation): Data from 42,386 volunteers, serving as a "low-risk" healthy volunteer cohort to test if findings held up in the general population.
Key Results
The AI identified sex with high accuracy, achieving an AUROC of 0.943 in the BIDMC cohort and 0.971 in the UK Biobank.
The "Male-Like" Female Risk
The most striking finding was that females with high discordance scores—ECGs that look more "male" to the AI—face a significantly higher risk of cardiovascular death. In female outpatients with clinically "normal" ECGs, an increased sex discordance score was associated with a 78% increased risk of cardiovascular death (Hazard Ratio 1.78). Interestingly, this relationship was not observed in males with "female-like" ECGs.
Biological Grounding
Phenome-wide (PheWAS) and Genome-wide (GWAS) association studies linked high discordance scores in females to increased left ventricular mass, reduced body fat, and genetic loci related to testosterone and insulin-like growth factors. This suggests the AI is picking up on real biological differences rather than just statistical noise.
Discussion and Limitations
The AI successfully identified that ECG phenotypes exist on a continuum. A "misclassification" by the AI is often a reflection of a biological phenotype that deviates from the sex-average, signaling underlying risk hidden in plain sight. While a human cardiologist might see a healthy heart, the AI detects subtle shifts in QRS voltage and T-wave morphology that signal vulnerability.
However, the study has limitations:
- Binary Constraints: Reliance on self-reported sex does not capture the complexities of gender identity or intersex conditions.
- Hormonal Data: Lack of concurrent blood hormone levels limits direct correlation with endocrine states at the time of ECG.
- Clinical Integration: Prospective trials are required to determine if flagging high-discordance females for aggressive intervention improves outcomes.
Integrating this AI-derived score into electronic health records could help close the "sex gap" in cardiology by identifying high-risk women who currently fly under the radar.
Summary updated