Natural & Applied Sciences — Bio-informatics Track

BrainState-ML: Deciphering Cognitive Assessment Performance and Brain Functional Organization from Task-fMRI Using Interpretable Machine Learning

Eason ZhouGlenview Senior Public School, Canada (Lead Author)
Dr. Haoyan JiangFutureEra Research Institute, Canada (Advisor)

An interpretable machine learning framework for decoding four distinct cognitive states from task-fMRI contrast maps using Schaefer 2018 cortical atlas parcellation ($100$ ROIs) and Yeo 7 functional network mapping.

85.83%
Random Forest Acc
0.9656
Logistic Reg AUROC
100 ROIs
Schaefer Atlas
120 Maps
Brainomics Dataset

fMRI Neuroimaging Brain Slicer & Statistical Maps

Interactive multi-plane slicing (Axial, Coronal, Sagittal) with Nilearn group statistical contrast maps.

Full Interactive Explorer

fMRI Neuroimaging Brain Visualizer

Switch between 3D Cortical Surface Mesh and 2D/3D Anatomical Slice Explorer across MNI coordinates.

Surface Model: FreeSurfer fsaverage5 (20k Vertices, 40k Faces)

5-Stage Methodology Pipeline

End-to-end task-fMRI decoding and network-level interpretation framework

Stage 01

Brainomics Dataset

30 subjects $\times$ 4 task contrasts = 120 statistical activation maps (Math, Reading, Motor, Visual).

Stage 02

Schaefer-100 ROIs

Extract mean activations per parcel into 100D feature vector X ∈ ℝ120×100.

Stage 03

Stratified 5-Fold CV

Grid search hyperparameter tuning across Logistic Regression, Random Forest, Decision Tree, and KNN.

Stage 04

Model Benchmarking

Random Forest achieved highest accuracy ($85.83\%$), Logistic Regression selected for interpretability ($0.9656$ AUROC).

Stage 05

Yeo-7 Interpretation

Mapped coefficient weights (Mk,n) to canonical functional networks (Somatomotor, Control, Visual, etc.).

Classifier Performance Benchmark (%)

Comparing 4 supervised models under Stratified 5-Fold Cross-Validation

N = 120 Contrast Maps

Yeo-7 Network Discriminative Importance (Mk,n)

Mean absolute coefficient weight mapped per functional network

Key Empirical Findings

Sensorimotor Separability

Motor Control and Visual Processing contrasts produced distinct, highly spatially localized activation signatures in somatomotor and occipital cortices, achieving near-perfect classification (1.000 AUROC).

Higher-Order Cognitive Overlap

Most classification errors occurred between Working Memory / Math and Language / Reading due to shared frontoparietal executive control and symbolic processing networks.

Functional Network Mapping

Frontoparietal Control and Dorsal Attention networks dominated Math decoding ($0.35$), while Default Mode Network contributed strongly to Language comprehension ($0.34$).