Classification of ROI-based fMRI data in short-term memory tasks using discriminant analysis and neural networks

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Understanding brain function requires analyzing spatiotemporal patterns in brain activity.
• Functional magnetic resonance imaging (fMRI) and machine learning (ML) are key tools for detecting functional connections between brain regions of interest (ROIs).
• Matching ML methods to specific problems is crucial for extracting meaningful information about ROI connections.

Purpose of the Study:

• To apply ML techniques for classifying tasks in a working memory experiment.
• To identify specific brain areas involved in information processing during working memory.
• To differentiate brain responses to visual stimuli (visuospatial, verbal) and experimental phases (encoding, retrieval).

Main Methods:

• Utilized classical discriminators and neural networks (convolutional, residual) for brain activity classification.
• Employed an algorithm considering feature correlations to identify important ROIs for model accuracy.
• Analyzed fMRI data to distinguish between resting state, encoding, and retrieval phases of working memory tasks.

Main Results:

• The LGBM classifier performed best with single-time point data during memory retrieval.
• A convolutional neural network achieved top performance during the encoding phase.
• Brain signals during retrieval were clearly distinguishable from resting state, unlike encoding signals.
• Confirmed the crucial role of basal ganglia in information processing during memory retrieval.

Conclusions:

• ML algorithms are beneficial for investigating working memory dynamics.
• Identified key brain regions and spatiotemporal distinctions in working memory encoding and retrieval processes.
• Working memory retrieval processes exhibit distinct brain signal structures compared to encoding and resting states.