Joint Data Harmonization and Group Cardinality Constrained Classification.

Area of Science:

• Neuroimaging
• Machine Learning
• Biostatistics

Background:

• Combining independent datasets enhances classifier power but requires data harmonization.
• Existing harmonization methods can reduce classification accuracy by mitigating group differences.
• Preserving distinct cohort characteristics is crucial for accurate disease classification.

Purpose of the Study:

• To propose a novel model unifying linear regression for data harmonization with logistic regression for disease classification.
• To develop an adaptive harmonization process that considers both disease and control data.
• To improve the accuracy of disease classification in neuroimaging studies using independent datasets.

Main Methods:

• A unified model combining linear regression for harmonization and logistic regression for classification.
• Adaptive data harmonization learning from both disease and control cohorts.
• Sparsity constraints (l0-norm) to reduce overfitting, account for regional inter-dependencies, and identify disease-specific patterns.

Main Results:

• The proposed model effectively distinguishes HIV-Associated Neurocognitive Disorder from Mild Cognitive Impairment.
• The classifier demonstrates impartiality to acquisition differences between independent datasets.
• Achieved higher accuracy in disease separation compared to sequential harmonization and classification methods and non-sparsity based logistic regression.

Conclusions:

• The unified, adaptive harmonization and classification model offers superior performance in neuroimaging studies.
• This approach enhances disease separation accuracy while being robust to multi-site data acquisition variations.
• The method provides a powerful tool for analyzing independent neuroimaging datasets for disease classification.