multiview-information-bottleneck-brain-hoi

name: multiview-information-bottleneck-brain-hoi description: "Multi-view Information Bottleneck framework for modeling higher-order interactions (HOIs) in resting-state fMRI for psychiatric diagnosis. Captures complex brain dynamics beyond pairwise connectivity without predefined hyperedges."

Multi-View Information Bottleneck for Higher-Order Brain Interactions

Information-theoretic approach for learning higher-order brain interactions from multi-view fMRI data using the Information Bottleneck principle.

Metadata

• Source: arXiv:2604.17713v1
• Title: Modeling Higher-Order Brain Interactions via a Multi-View Information Bottleneck Framework for fMRI-based Psychiatric Diagnosis
• Authors: Kunyu Zhang, Qiang Li, Vince D. Calhoun, et al.
• Published: 2026-04-20
• Category: Neuroscience/fMRI Analysis

Core Methodology

Problem Context

Resting-state fMRI is crucial for psychiatric diagnosis, but most approaches rely on pairwise connectivities that overlook higher-order interactions (HOIs) central to complex brain dynamics. Hypergraph methods require predefined hyperedges.

Information Bottleneck Solution

Multi-View IB Framework:

1. Information Bottleneck principle: Trade-off between compression and prediction
2. Multi-view learning: Multiple perspectives on brain data
3. HOI discovery: Learn higher-order interactions from data
4. No predefined hyperedges: Flexible structure learning

Key Innovation

• IB-based HOI modeling: Information-theoretic higher-order interaction discovery
• Multi-view integration: Combines multiple fMRI representations
• Flexible structure: No fixed hypergraph topology required
• Psychiatric diagnosis: Application to mental health classification

Technical Framework

Information Bottleneck Principle

Objective: max I(T; Y) - β·I(T; X)

Where:
- X: Input fMRI data (multi-view)
- T: Compressed representation (bottleneck)
- Y: Target label (diagnosis)
- β: Compression-prediction trade-off

Multi-View Architecture

fMRI Data View 1 ──┐
fMRI Data View 2 ──┼──> Information Bottleneck ──> HOI Representation ──> Classification
fMRI Data View 3 ──┘

Higher-Order Interaction Learning

• Synergistic information: Information in joint distribution
• Redundant information: Shared across views
• Unique information: View-specific contributions
• HOI extraction: Beyond pairwise correlations

Implementation Guide

Prerequisites

• Resting-state fMRI data
• Multiple preprocessing pipelines (for multi-view)
• Python with PyTorch/TensorFlow
• nilearn, scikit-learn for neuroimaging

Steps

1. Data preparation: Multi-view fMRI preprocessing
2. View creation: Different atlases, bandpass filters, parcellations
3. IB optimization: Train bottleneck representation
4. HOI extraction: Compute higher-order interaction strengths
5. Classification: Psychiatric diagnosis using HOI features

Code Structure

import torch
from information_bottleneck import MultiViewIB

# Initialize multi-view IB
mib = MultiViewIB(
    n_views=3,
    input_dims=[116, 200, 400],  # Different parcellations
    hidden_dim=128,
    beta=0.5
)

# Multi-view fMRI data
views = [fmri_view1, fmri_view2, fmri_view3]

# Forward pass
bottleneck, hoi_features = mib(views)

# Classification
logits = classifier(hoi_features)
loss = classification_loss + mib.information_loss

# HOI analysis
hoi_strength = mib.extract_hoi_interactions()

Applications

• Psychiatric diagnosis (depression, schizophrenia, ADHD)
• Brain connectivity analysis
• Higher-order interaction discovery
• Multi-modal neuroimaging fusion
• Mental health biomarker discovery

Information-Theoretic Metrics

• Multi-information: Total correlation in HOI
• Synergy: Information gain from joint consideration
• Redundancy: Overlapping information across views
• Integration: Measures of HOI strength

Related Skills

• multi-view-o-information-brain-dynamics
• multi-view-o-information-brain-hoi
• functional-connectivity-graph-neural-networks

References

• arXiv:2604.17713v1
• Tishby et al. Information Bottleneck method