A Reproducible Framework for Bias-Resistant Machine Learning on Small-Sample Neuroimaging Data.

Area of Science:

• Neuroimaging
• Machine Learning
• Biomedical Data Analysis

Background:

• Conventional cross-validation methods can lead to biased performance estimates in machine learning models.
• This bias hinders reproducibility and generalization, especially with limited data.
• Small-sample neuroimaging datasets present unique challenges for reliable model development.

Purpose of the Study:

• To introduce a reproducible and bias-resistant machine learning framework for small-sample neuroimaging data.
• To address limitations in conventional cross-validation for model selection and performance estimation.
• To provide a generalizable computational blueprint for reliable machine learning in data-limited biomedical domains.

Main Methods:

• Integration of domain-informed feature engineering.
• Implementation of nested cross-validation for unbiased performance estimation.
• Calibrated decision-threshold optimization for robust classification.

Main Results:

• The framework achieved a nested cross-validation balanced accuracy of 0.660 ± 0.068 on a structural MRI dataset.
• A compact, interpretable subset of features was selected using importance-guided ranking.
• Demonstrated improved reliability and interpretability compared to conventional methods.

Conclusions:

• The developed framework offers a reproducible and bias-resistant approach for machine learning in neuroimaging.
• It enables reliable evaluation and feature selection for small-sample biomedical datasets.
• This work provides a computational blueprint for advancing machine learning applications in data-limited research areas.