Abstract 4141659: Uncovering Sex Bias in Machine Learning Algorithms for Detecting Acute Myocardial Infarction using Electrocardiographic Data

Introduction: Acute myocardial infarction (AMI) is a life-threatening manifestation of cardiovascular disease so timely detection and intervention are crucial for improving patient outcomes. Increasingly, machine learning (ML) algorithms are being leveraged to analyze 12-lead electrocardiograms (ECGs) to optimize early and accurate triage for AMI. However, the potential inherent sex bias within these data-driven algorithms remains insufficiently explored. This study aims to uncover potential sex bias in ML algorithms through a systematic and well-controlled investigation. Methods: Utilizing a public ECG dataset (PTBXL) with cardiologist annotations of AMI and non-AMI classes, we conducted sex-based resampling to obtain datasets with varying male-to-female sample ratio from 0.5 to 1.5, keeping the AMI to non-AMI ratio constant at 1:2 for each sex group. We evaluated the performance of ML models derived from these datasets on ECGs from male and female subgroups. The process was repeated 30 times with random subsampling. The area under the precision-recall curve (AUCPR) was used as the performance metric because it is robust to class imbalance issues. Results: The models consistently performed better on the male subgroup, even as the number of female samples increased. At a male-to-female ratio of 0.5 (male: 2394; female: 4788), the mean AUCPR for males was 0.626 ± 0.03, compared to 0.573 ± 0.02 for females. This trend persisted at a balanced ratio of 1.0 (4788 samples each), with an AUCPR of 0.651 ± 0.02 for males and 0.592 ± 0.02 for females. At a ratio of 1.5 (male: 7182; female: 4788), the AUCPR for males was 0.667 ± 0.02, versus 0.598 ± 0.02 for females. All comparisons were statistically significant (p < 0.05, two-sample Student t-tests). Discussion: Despite an increased number of female samples, the consistently better performance of the model to detect AMI in males suggests a potential bias in the model. We hypothesize that this bias may be related to inherent differences in the distribution of ECG patterns between male and female subgroups. Our findings indicate that simply balancing the dataset by increasing the female sample size is insufficient to mitigate bias. Addressing disparities in healthcare data requires more nuanced strategies beyond data balancing. Future work should explore alternative approaches to overcome sex bias in the application of ML algorithms in the interpretation of ECGs in patients with suspected AMI.