Analyzing neuroimaging data with subclasses: A shrinkage approach.

Area of Science:

• Neuroimaging analysis
• Machine learning for neuroscience

Background:

• Linear Discriminant Analysis (LDA) is a common method for binary classification in neuroimaging.
• Standard LDA is suboptimal when subclass labels, providing valuable data structure, are available.
• Existing methods fail to effectively leverage subclass information for improved classification.

Purpose of the Study:

• To introduce a novel method, Relevance Subclass LDA (RSLDA), for incorporating subclass labels into neuroimaging data classification.
• To demonstrate the effectiveness of RSLDA in enhancing classification accuracy and data interpretation.
• To validate RSLDA on diverse neuroimaging datasets, including EEG and fMRI.

Main Methods:

• Developed RSLDA, a novel classifier that computes individual hyperplanes for each subclass.
• Employed regularized estimators of subclass means, using other subclasses for regularization.
• Applied RSLDA to electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) datasets.

Main Results:

• RSLDA significantly outperformed standard LDA on both EEG and fMRI datasets.
• The method successfully exploited subclass structure, leading to increased classification accuracy.
• RSLDA generated regularization profiles, offering meaningful interpretation of subclass structure.

Conclusions:

• RSLDA offers a significant advancement for binary classification in neuroimaging by effectively utilizing subclass labels.
• The method provides both enhanced classification performance and improved interpretability of neuroimaging data.
• RSLDA is recommended for a wide range of classification tasks within and beyond the field of neuroimaging.