Enhancing Fairness in Disease Prediction by Optimizing Multiple Domain Adversarial Networks

Area of Science:

• Biomedical informatics
• Machine learning in healthcare
• Artificial intelligence for disease prediction

Background:

• Biomedical predictive models require equitable and reliable outcomes.
• Algorithmic bias in medical predictions exacerbates health disparities.
• Addressing bias is crucial for fair and effective healthcare applications.

Approach:

• Introduced a Multiple Domain Adversarial Neural Network (MDANN) framework.
• Utilized multiple adversarial components for learning fair patterns via negative gradient back-propagation across sensitive features.
• Employed Area Under the ROC Curve (AUC) loss functions to manage class imbalance and enhance minority group performance.
• Integrated pre-trained convolutional autoencoders (CAEs) for deep data representation to boost accuracy and fairness.

Key Points:

• MDANN framework effectively mitigates biases and disparities in disease prediction.
• AUC loss functions promote equitable classification performance, especially for underrepresented populations.
• CAEs enhance predictive accuracy and fairness by extracting robust data representations.
• Empirical results show superior accuracy and fairness compared to state-of-the-art methods.

Conclusions:

• The MDANN approach provides reliable and equitable disease prediction.
• Demonstrated improved accuracy and fairness in predicting Alzheimer's and Autism progression using brain imaging data.
• Highlights the potential of adversarial learning and deep representation for unbiased biomedical AI.