Generative embedding for model-based classification of fMRI data.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Multivariate classification algorithms using fMRI data are common for inferring brain states.
• Current methods face challenges in high-dimensional, low-sample data, leading to poor generalization.
• Discriminative methods like SVMs often lack mechanistic interpretability.

Purpose of the Study:

• To propose a novel generative-embedding approach for fMRI data analysis.
• To improve classification accuracy and mechanistic interpretability in brain state decoding.
• To extend trial-by-trial classification to subject-by-subject classification for fMRI.

Main Methods:

• Developed a generative-embedding approach combining Dynamic Causal Models (DCMs) with Support Vector Machines (SVMs).
• Proposed a general procedure for DCM-based generative embedding for subject-wise classification in fMRI.
• Illustrated the method with a clinical example classifying aphasic patients and healthy controls using DCM of thalamo-temporal regions.

Main Results:

• The generative embedding approach achieved a near-perfect balanced classification accuracy of 98% in the clinical example.
• Outperformed conventional activation-based and correlation-based methods significantly.
• Demonstrated detection of disease states with high accuracy and mechanistic interpretation of connectivity abnormalities.

Conclusions:

• Generative embedding offers a powerful tool for accurate and interpretable brain state classification using fMRI.
• This method can reveal mechanistic insights into clinical conditions by analyzing connectivity.
• Future applications may advance the subtyping of spectrum disorders into more physiologically defined groups.