Mitigating Algorithmic Bias in AI-Driven Cardiovascular Imaging for Fairer Diagnostics

Area of Science:

• Artificial Intelligence
• Cardiovascular Medicine
• Medical Imaging

Background:

• Algorithmic bias in deep learning models for cardiovascular risk prediction exacerbates health disparities.
• Existing models lack fairness across demographic and socioeconomic groups, necessitating bias mitigation strategies.
• Integration of fairness-aware algorithms and explainable AI is crucial for equitable healthcare.

Purpose of the Study:

• To address algorithmic bias in deep learning for cardiovascular risk prediction.
• To integrate fairness-aware algorithms and susceptible carrier-infected-recovered (SCIR) models for equitable AI insights.
• To enhance diagnostic precision and reduce disparities in vulnerable populations.

Main Methods:

• Utilized 3D/4D cardiac MRI and tabular data from the Cardiac Atlas Project (CAP).
• Adapted the SCIR model with adversarial debiasing, Fairlearn, and equalized odds post-processing.
• Implemented YOLOv5, Mask R-CNN, and ResNet18 with LIME and SHAP for interpretability.

Main Results:

• Bias mitigation improved disparate impact (0.80 to 0.95) and reduced equal opportunity difference (0.20 to 0.05).
• SCIR model demonstrated robustness with high Intersection over Union (IoU) scores (94.8%) and Dice coefficients (0.941-0.980).
• Interpretability models (YOLOv5, Mask R-CNN, ResNet) showed strong performance in segmentation and classification.

Conclusions:

• Fairness-aware algorithms effectively address biases in cardiovascular predictive models.
• Integration of fairness and explainable AI promotes equitable diagnostic precision.
• The study reduces diagnostic disparities, enhancing clinical trust in AI systems.