Statistically valid explainable black-box machine learning: applications in sex classification across species using...

Area of Science:

• Neuroimaging
• Machine Learning
• Comparative Anatomy

Background:

• Sex classification from neuroimaging data offers potential for personalized diagnostics by identifying sex-specific brain differences linked to disease risks.
• Traditional machine learning methods struggle with high-dimensional neuroimaging data, lacking accuracy and interpretable feature importance.
• Existing techniques like random forests, LIME, and SHAP face challenges with complex feature interactions and noise in large datasets.

Purpose of the Study:

• To develop an integrated framework combining Oblique Random Forests (ORFs) and a novel permutation-based feature importance testing algorithm (NEOFIT).
• To enhance classification accuracy and provide statistically validated, interpretable feature importance for sex classification using neuroimaging data.
• To enable cross-species comparison of sex-based brain structure differences in humans and macaques.

Main Methods:

• Developed an integrated framework using Oblique Random Forests (ORFs) with oblique decision boundaries for complex feature interactions.
• Introduced NEOFIT, a permutation-based feature importance testing algorithm for rigorous statistical validation and corrected p-values.
• Validated the framework on simulated datasets and applied it to voxel-wise structural MRI and cortical thickness data in humans and macaques.

Main Results:

• ORFs achieved high classification accuracy, with AUC > 0.80 in humans and > 0.70 in macaques.
• NEOFIT identified statistically significant neuroanatomical features aligned with sex-dimorphic patterns.
• The framework demonstrated robustness, scalability, and provided interpretable insights into sex-distinguishing brain features.

Conclusions:

• The proposed framework significantly enhances sex classification performance from neuroimaging data.
• It provides clear, interpretable insights into neuroanatomical features that differentiate sexes, aiding in understanding sex-specific disease risks.
• These advancements contribute to improved diagnostic tools and a deeper understanding of the evolutionary basis of sex differences in brain structure.