Related Experiment Video
Updated: Jun 6, 2026
14:27
Identification of Disease-related Spatial Covariance Patterns using Neuroimaging Data
Published on: June 26, 2013
Supervised Domain Adaptation Mitigates Cross-Ethnicity Prediction Error in Neuroimaging-Based Cognitive Prediction
Farzane Lal Khakpoor1, William van der Vliet1, Jeremiah Deng2
1Department of Psychology, University of Otago, Dunedin, New Zealand.
Biorxiv : the Preprint Server for Biology
|June 5, 2026
Summary
Supervised domain adaptation methods, like balanced weighting, reduce racial bias in neuroimaging AI. These techniques improve model fairness and generalizability across diverse populations, particularly for structural MRI predictions.
Area of Science:
- Neuroimaging
- Machine Learning
- Artificial Intelligence
Background:
- Machine learning models predict cognitive/clinical outcomes from neuroimaging data.
- Racial/ethnic imbalances in datasets cause performance disparities in AI models.
- Models often perform better for majority groups, limiting generalizability.
Purpose of the Study:
- Evaluate supervised domain adaptation methods to mitigate bias in neuroimaging AI.
- Assess if methods improve model generalization from White American to African American participants.
- Identify effective strategies for equitable AI in neuroimaging.
Main Methods:
- Used the ABCD dataset with 80 MRI measures.
- Applied domain adaptation techniques: balanced weighting, TrAdaBoost, SrcOnly, linear interpolation.
- Trained models on White American participants and tested generalization to African American participants.
Main Results:
- All domain adaptation methods reduced prediction error for African American participants.
- Balanced weighting showed the best performance and stability, even with limited data.
- Improvements were most significant for modalities with large baseline disparities, like structural MRI.
Conclusions:
- Simple, low-cost domain adaptation strategies can effectively reduce cross-ethnic performance gaps.
- Balanced weighting offers a practical approach to enhance equity in predictive neuroimaging.
- Findings support the development of more equitable neuroimaging predictive biomarkers.