Generative AI mitigates representation bias and improves model fairness through synthetic health data

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Biostatistics

Background:

• Representation bias in health data leads to unfair clinical decisions and limits research generalizability.
• Underrepresented groups, including specific ethnic and gender populations, do not equally benefit from clinical advancements.
• Existing methods for mitigating bias, like SMOTE and Generative Adversarial Networks (GANs), struggle with high-dimensional time-series health data.

Purpose of the Study:

• To develop a novel architecture, CA-GAN, capable of synthesizing authentic, high-dimensional time-series health data.
• To address the challenge of generating realistic synthetic data for underrepresented subpopulations.
• To improve fairness and performance of AI models in clinical settings.

Main Methods:

• Developed a novel Conditional Attention Generative Adversarial Network (CA-GAN) architecture.
• Utilized two diverse, real-world clinical datasets comprising 7535 patients with hypotension and sepsis.
• Evaluated CA-GAN against state-of-the-art methods using qualitative and quantitative metrics, including assessment for mode collapse.

Main Results:

• CA-GAN successfully synthesizes authentic, high-dimensional time-series health data, outperforming existing methods.
• The generated synthetic data demonstrably improves model fairness for underrepresented groups, specifically Black patients and female patients.
• CA-GAN effectively generates minority class data while preserving the original data distribution, leading to enhanced downstream predictive task performance.

Conclusions:

• CA-GAN offers a robust solution for generating high-quality synthetic health data, mitigating representation bias.
• The proposed method enhances fairness and generalizability of AI models in healthcare, benefiting underrepresented populations.
• CA-GAN represents a significant advancement in addressing data bias for improved clinical AI applications.