Fairness-related performance and explainability effects in deep learning models for brain image analysis.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Ethics

Background:

• Explainability and fairness are crucial for ethical AI in healthcare.
• Limited research exists on how unfairness appears in explainable AI (XAI) and its use in investigating bias.
• Sociodemographic factors can act as confounders in machine learning models.

Purpose of the Study:

• To analyze the impact of sociodemographic confounders on classifier performance and XAI methods.
• To investigate performance disparities in deep learning models across different demographic groups.
• To explore how XAI can reveal underlying reasons for unfairness in AI predictions.

Main Methods:

• A convolutional neural network (CNN) was trained on T1-weighted brain MRI data from 4547 adolescents (9-10 years old).
• The CNN predicted biological sex, and performance was analyzed for disparities between White and Black subjects.
• Saliency maps were generated for subgroups (sex and race intersection) to visualize model attention.

Main Results:

• The CNN exhibited significant performance differences in classifying males between White and Black adolescents.
• Slightly higher performance was observed for Black females compared to White females.
• Saliency maps revealed subgroup-specific patterns in brain regions associated with pubertal development, a race-linked confounder.

Conclusions:

• A CNN showed performance disparities in sex classification between Black and White adolescents.
• Subgroup saliency maps identified different important brain regions, linked to pubertal development differences.
• This study demonstrates that unfair AI models can yield varying XAI results, offering insights into bias origins.