Sex classification from functional brain connectivity: Generalization to multiple datasets.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Machine learning (ML) models in neuroscience often face limited training data, hindering their generalizability.
• Optimizing training sample characteristics for robust ML model generalization remains an open question.

Purpose of the Study:

• To systematically evaluate the impact of training sample composition on the generalization performance of ML-based sex classification models using neuroimaging data.
• To determine whether single or compound samples, and sample size, best enhance classifier generalizability.

Main Methods:

• Developed parcelwise classifiers (pwCs) for sex classification based on neuroimaging connectivity profiles.
• Compared generalization performance of pwCs trained on single dataset samples versus compound samples of varying sizes.
• Quantified generalization using mean across-sample classification accuracy and spatial consistency.

Main Results:

• Generalization performance of pwCs trained on single samples varied depending on the specific test datasets.
• Compound samples consistently yielded the highest generalization performance across all test datasets.
• Models trained on compound samples generalized well even to datasets not included in the training set.

Conclusions:

• Both large sample size and heterogeneous data composition are crucial for achieving generalizable ML models in neuroimaging.
• Training ML models on diverse, combined datasets enhances their ability to perform accurately on unseen data.