Adversarial Debiasing for Equitable and Fair Detection of Acute Coronary Syndrome Using 12-Lead ECG

Area of Science:

• Cardiology
• Artificial Intelligence
• Health Equity

Background:

• Accurate diagnosis of acute coronary syndrome (ACS) is critical for patient outcomes.
• Racial disparities in symptom presentation can lead to diagnostic inaccuracies.
• Machine learning offers potential for improving ACS diagnosis and reducing disparities.

Purpose of the Study:

• To develop and evaluate machine learning models for diagnosing ACS.
• To mitigate diagnostic disparities between Black and non-Black populations.
• To ensure fairness in diagnostic accuracy across racial subgroups.

Main Methods:

• Utilized a random forest classifier framework.
• Compared data resampling, data partitioning, and adversarial debiasing strategies.
• Evaluated model performance using receiver operating characteristic (ROC) curves and sensitivity at 80% specificity.

Main Results:

• Adversarial debiasing reduced the sensitivity difference between racial subgroups from 9.8% to 1.3%.
• The adversarial debiasing model achieved areas under the ROC of 0.810 (Black) and 0.817 (non-Black).
• Sensitivities were 70.1% for Black and 71.4% for non-Black subgroups post-mitigation.

Conclusions:

• Adversarial debiasing demonstrated superior performance in both diagnostic accuracy and disparity reduction compared to other methods.
• This framework is expected to enable fair diagnostic models for diverse global populations and be applicable to other health outcomes.