Using Copula distributions to support more accurate imaging-based diagnostic classifiers for neuropsychiatric...

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Psychiatry

Background:

• Machine learning (ML) applied to brain magnetic resonance images (MRIs) for neuropsychiatric disorder diagnosis faces challenges due to high-dimensional data from limited participants.
• Sparse data in high-dimensional spaces increase classifier variability, limiting validity, reproducibility, and generalizability.
• Accurate estimation of multivariate distributions of imaging measures is crucial for improving classifier accuracy and stability.

Purpose of the Study:

• To develop a novel method for accurately estimating multivariate distributions of brain imaging measures from sparse data.
• To improve the accuracy, reproducibility, validity, and generalizability of machine learning classifiers for neuropsychiatric disorder diagnosis.
• To leverage Copula methods for generating dense imaging datasets to train more robust classifiers.

Main Methods:

• Proposed a method to estimate univariate distributions of imaging data first.
• Combined univariate distributions using a Copula to generate accurate multivariate distribution estimates.
• Sampled Copula distributions to create dense sets of imaging measures for training machine learning classifiers.

Main Results:

• Classifiers trained on imaging measures sampled from Copula distributions demonstrated significantly higher accuracy and reproducibility.
• Copula-based classifiers outperformed those trained on real-world imaging measures or multivariate Gaussian distributions.
• The method effectively addressed the challenges of sparse, high-dimensional neuroimaging data.

Conclusions:

• Estimated multivariate Copula distributions provide a robust way to generate dense brain imaging datasets.
• This approach significantly enhances the accuracy and reproducibility of diagnostic classification algorithms in clinical neuroimaging.
• The Copula-based method offers a promising solution for improving ML-driven diagnosis of neuropsychiatric disorders.