Shift happens: a fairness-oriented framework for medical classification under hidden bias.

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Medical Data Analysis

Background:

• Medical AI models often exhibit performance disparities across patient groups due to biased training data.
• Hidden biases in AI can lead to unreliable and unfair clinical decision-making.
• Ensuring fairness and equity in AI-driven healthcare is a critical challenge.

Purpose of the Study:

• To develop an AI system, DPE-Former, that maintains accuracy and fairness across diverse patient subpopulations, irrespective of explicit labeling.
• To mitigate performance gaps in medical AI caused by biased data.
• To enhance the reliability and equity of AI tools in clinical settings.

Main Methods:

• Introduction of DPE-Former, a model integrating prototype-based learning with transformer attention.
• Training complementary classifiers on balanced data subsets to capture diverse population aspects.
• Utilizing a transformer module for adaptive output combination to balance decisions for unseen or minority groups.

Main Results:

• DPE-Former demonstrated superior accuracy on underrepresented groups across diverse datasets (prostate ultrasound, skin lesions, cardiac records).
• The model achieved more consistent performance compared to standard training methods.
• Evidence of reduced bias and improved fairness in AI model predictions.

Conclusions:

• DPE-Former presents a straightforward yet potent method for minimizing hidden bias in medical AI.
• The system enhances fairness and reliability for both image and tabular medical data.
• Supports equitable decision-making in critical clinical applications like cancer diagnosis and cardiac care.