Improving medical machine learning models with generative balancing for equity and excellence

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Health Equity

Background:

• Machine learning for clinical outcome prediction faces challenges with imbalanced datasets and persistent biases, leading to health inequities.
• Existing methods struggle to address rare outcomes and ensure equitable treatment across diverse patient populations.
• Disparities in data representation and scarcity of positive labels exacerbate prediction biases.

Purpose of the Study:

• To introduce FairPlay, a novel synthetic data generation approach using large language models.
• To enhance algorithmic performance and mitigate bias in clinical outcome prediction.
• To improve representation and preserve patient privacy in healthcare datasets.

Main Methods:

• Leveraging large language models (LLMs) to generate realistic, anonymous synthetic patient data.
• Augmenting existing datasets to improve representation and capture complex patterns.
• Utilizing synthetic data to train and evaluate machine learning models for clinical outcome prediction.

Main Results:

• FairPlay significantly boosts mortality prediction performance across diverse patient subgroups, with up to a 21% F1 Score improvement.
• The approach reduces subgroup performance gaps without requiring additional real-world data or altering training pipelines.
• Consistent improvements in both performance and fairness metrics were observed across multiple experimental setups.

Conclusions:

• FairPlay offers an effective strategy for generating privacy-preserving synthetic data to address bias and performance limitations in clinical ML.
• This method enhances the fairness and accuracy of predictive models, contributing to more equitable healthcare.
• Synthetic data generation with LLMs presents a promising avenue for improving AI applications in sensitive medical domains.