Diagnosing mental disorders such as autism, depression, or schizophrenia goes beyond taking simple brain images. Resting-state fMRI (rs-fMRI) observes brain activity while at rest, revealing which regions activate simultaneously. This forms the basis for functional connectivity.
Traditional studies have used graphs and neural networks, but they mostly focus on pairwise interactions — asking “do regions A and B co-activate?” But what about higher-order relationships, like among regions A, B, and C all at once?
This is exactly where MvHo-IB comes in — a method that combines deep learning, theoretical information science, and brain data.
🔍 What is MvHo-IB?
🧠 Multi-View = Two Perspectives on the Data
MvHo-IB stands for Multi-view Higher-order Information Bottleneck. It analyzes brain data from two views:
- View 1 – Pairwise FC: traditional correlation matrix of brain region activity pairs.
- View 2 – HOI Tensor: a 3D representation of triplet-wise interactions, derived from fMRI using O-information.
📦 What is the Information Bottleneck?
According to classical information theory (Tishby et al., 2000), a model should:
preserve only the information from input $X$ that is relevant to predicting output $Y$ (e.g., the diagnosis).
Formally:
$$ \max_{T} ; I(T; Y) - \beta I(T; X) $$
Where:
- $T$: a compressed representation (e.g. a hidden layer in a neural net),
- $I(\cdot;\cdot)$: mutual information,
- $\beta$: balances usefulness vs simplicity.
MvHo-IB extends this idea to higher-order interactions in the brain.
🧮 O‑Information: The Missing Piece
Standard mutual information handles only two variables. But what about groups — are regions A, B, and C forming a coordinated pattern?
That’s where O‑information (Rosas et al., 2020) comes in — a measure that reveals whether multi-variable interactions are:
- Redundant (repeating the same info),
- Synergistic (together revealing more than parts alone).
This allows us to identify groups of brain regions that:
- may not be individually informative,
- but collectively signal mental disorder.
🧰 MvHo‑IB Architecture
Extracting fMRI features:
- Pairwise FC: standard correlation matrix (e.g., 116 × 116),
- HOI: 3D tensor (116 × 116 × 116) containing triplet O‑information.
Brain3D‑CNN:
- A custom 3D convolutional neural network designed to process HOI as volumetric brain data.
Multi-view Information Bottleneck Loss:
- A loss function that keeps only the features relevant to diagnosis, discarding noise.
📈 Experiments and Results
Datasets:
- ABIDE, ADHD-200, and COBRE — well-known rs-fMRI benchmarks,
- Hundreds of patients with both diagnosed and control groups.
Results:
- MvHo-IB outperformed all compared methods, including graphs, hypergraphs, and GCNs,
- Achieved up to +7% higher accuracy than the best existing methods.
🧭 What This Paper Contributes
| Contribution | Impact |
|---|---|
| 🎯 HOI + Pairwise FC | Higher-dimensional view of brain connectivity |
| 🧠 O-information | Captures complex, synergistic brain interactions |
| 🧰 Brain3D-CNN + IB Loss | Novel deep learning framework for fMRI |
| 📊 Real-world diagnostic improvements | More accurate classification of brain disorders |
📌 Summary
MvHo‑IB is a breakthrough method showing that thinking in higher-order groups (triplets or more) reveals critical patterns in the brain. By combining O‑information, 3D CNNs, and information bottleneck theory, it brings us closer to personalized, accurate diagnosis of mental health conditions.
📎 Links
- Based on the publication 📄 2507.02847