Advancing Fair and Explainable Machine Learning for Neuroimaging Dementia Pattern Classification in Multi-Ethnic...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Dementia affects millions globally, with diagnoses projected to triple by 2050.
• Accurate dementia diagnosis is crucial for treatment and quality of life.
• Current diagnostic tools show inconsistent precision and impartiality across diverse cultural groups.

Purpose of the Study:

• To investigate performance discrepancies in dementia classification among White American, African American, and Hispanic populations.
• To address cross-group bias in dementia diagnostic models.
• To introduce and evaluate novel techniques for improving model adaptability in underrepresented populations.

Main Methods:

• Investigated dementia classification performance across White American, African American, and Hispanic populations.
• Developed and applied a novel combination of few-shot learning and domain alignment.
• Assessed model adaptability and inter-group performance gaps.

Main Results:

• Significant cross-group bias was observed in dementia classification models, especially when models trained on one group were tested on another.
• The novel techniques substantially reduced inter-group performance gaps.
• Performance gaps were particularly reduced between White American and Hispanic cohorts.

Conclusions:

• Fairness-aware strategies and diverse training data are crucial for accurate and equitable dementia diagnoses.
• Few-shot learning and domain alignment show promise in mitigating bias in AI-driven medical diagnostics.
• Addressing disparities in dementia diagnosis is essential for improving patient outcomes across all populations.