Boltzmann-Inspired Model for fMRI Time-Series Classification

Joseph Van Duyn, Pavel Popov, Armin Iraji, Zening Fu, Vince D. Calhoun, Sergey Plis, Alex Fedorov

Keywords: resting-state fMRI, machine learning, deep learning, brain disorders, predictive neuroimaging

Track: Findings

Abstract: Accurate classification of psychiatric and neurological disorders using resting-state fMRI data is critical for understanding the mechanisms underlying brain diseases, yet existing approaches often rely on complex architectures that are difficult to interpret. In this work, we introduce a Boltzmann-inspired Model (BiM) for fMRI classification that leverages bilinear interactions between features within a lightweight mean aggregation framework. Our model captures higher-order dependencies among brain regions. We evaluate the proposed model across multiple brain disorders data and parcellations based on Independent Component Analysis (ICA) and Regions of Interest (ROIs). The performance of the proposed model is compared with state-of-the-art time series and functional network connectivity models on the classification of diseases and disorders. The proposed model demonstrates competitive or superior Area Under the Receiver Operating Characteristic Curve (ROC AUC) performance, particularly using ROI-based features. These results suggest that relatively simple, bilinear architectures can match or surpass more complex models.

General Area: Models and Methods

Specific Subject Areas: Time Series, Medical Imaging, Supervised Learning

Data And Code Availability: No

Ethics Board Approval: No

Entered Conflicts: I confirm the above

Anonymity: I confirm the above

Submission Number: 220