How do deep-learning models generalize across populations? Cross-ethnicity generalization of COPD detection

Area of Science:

• Artificial Intelligence in Medical Imaging
• Pulmonology
• Radiology

Background:

• Deep learning models for chronic obstructive pulmonary disease (COPD) detection on chest CT scans may exhibit performance disparities across different ethnic populations.
• Evaluating and mitigating potential biases in AI models is crucial for equitable healthcare delivery.

Purpose of the Study:

• To assess the performance and ethnic biases of deep learning models for COPD detection on chest CT scans.
• To compare supervised learning (SL) and self-supervised learning (SSL) strategies.
• To investigate the impact of training dataset composition (ethnic-specific vs. balanced) on model performance and bias.

Main Methods:

• Retrospective analysis of chest CT scans and clinical data from 7549 individuals (5240 non-Hispanic White, 2309 African American).
• Models were trained using different strategies: population-specific (NHW-only, AA-only), balanced (NHW + AA), and combined datasets.
• Comparison of supervised learning (SL) versus self-supervised learning (SSL) methods, including SimCLR, with assessment of distribution shifts across ethnicities.

Main Results:

• Self-supervised learning (SSL) methods significantly outperformed supervised learning (SL) methods (p < 0.001) across all training configurations.
• Training on balanced datasets comprising both non-Hispanic White and African American individuals improved model performance compared to population-specific training.
• Balanced datasets led to fewer distribution shifts between ethnic groups and health statuses, effectively reducing model bias.

Conclusions:

• Utilizing SSL methods and training on large, balanced, and diverse datasets enhances COPD detection model performance.
• This approach is effective in reducing performance and detection biases across different ethnic populations.
• Equitable AI-driven healthcare solutions for COPD diagnosis necessitate the use of diverse training data and advanced learning strategies.