BrainIB++: Leveraging Graph Neural Networks and Information Bottleneck for Functional Brain Biomarkers in Schizophrenia

TL;DR

BrainIB++ tackles the challenge of deriving interpretable brain biomarkers for schizophrenia from rs-fMRI by marrying graph neural networks with the information bottleneck. The method shifts to node-centric subgraph sampling guided by group-ICA parcellation, and optimizes a mutual-information objective to identify informative brain regions that drive classification. Across three large, multi-cohort datasets, BrainIB++ achieves superior accuracy and demonstrates robust generalization, while its subgraphs align with known clinical biomarkers in visual, sensorimotor, and higher-cognition networks. This work advances clinically meaningful, explainable brain biomarkers and offers a scalable framework for cross-site schizophrenia diagnosis with potential multi-modal extensions.

Abstract

The development of diagnostic models is gaining traction in the field of psychiatric disorders. Recently, machine learning classifiers based on resting-state functional magnetic resonance imaging (rs-fMRI) have been developed to identify brain biomarkers that differentiate psychiatric disorders from healthy controls. However, conventional machine learning-based diagnostic models often depend on extensive feature engineering, which introduces bias through manual intervention. While deep learning models are expected to operate without manual involvement, their lack of interpretability poses significant challenges in obtaining explainable and reliable brain biomarkers to support diagnostic decisions, ultimately limiting their clinical applicability. In this study, we introduce an end-to-end innovative graph neural network framework named BrainIB++, which applies the information bottleneck (IB) principle to identify the most informative data-driven brain regions as subgraphs during model training for interpretation. We evaluate the performance of our model against nine established brain network classification methods across three multi-cohort schizophrenia datasets. It consistently demonstrates superior diagnostic accuracy and exhibits generalizability to unseen data. Furthermore, the subgraphs identified by our model also correspond with established clinical biomarkers in schizophrenia, particularly emphasizing abnormalities in the visual, sensorimotor, and higher cognition brain functional network. This alignment enhances the model's interpretability and underscores its relevance for real-world diagnostic applications.

Figures (4)

• Figure 1: The pipeline to construct brain networks for graph neural networks from resting-state fMRI data. Data preprocessing of fMRI data involved preprocessing the raw resting-state fMRI data, followed by the application of spatially constrained group ICA using the NeuroMark_fMRI_2.1 template, which includes 105 ICNs to extract the corresponding time courses. The FNC matrices are computed using Pearson correlation between ICNs. From the FNC matrices, we construct the brain functional graph $G=\{X, A\}$. Specifically, $A$ is the adjacency matrix, where strong weighted functional connectivity values are binarized into ones between node pairs ($A \in \{0,1\}^{n \times n}$), while weaker connections are set to zero. $X$ is the node feature matrix with weighted functional connectivity values $(X \in \mathbb{R}^{n \times n})$. The weight feature of the $k$-th node $X_k$ defined as $X_k = [\rho_{k1}, \rho_{k2},...,\rho_{kn}]^T$, where $n$ is the number of the nodes, and $\rho_{kl}$ represents Pearson correlation coefficient between node $k$ and node $l$.
• Figure 2: Architecture of our proposed BrainIB++. BrainIB++ consists of three modules including a subgraph generator, a graph encoder, and a mutual information estimator. Given the input graph data G, the subgraph generator samples subgraphs $G_{\text{sub}}$ along with the node assignment, indicating whether a specific node belongs to ${G}_{\text{sub}}$ or $\overline{{G}_{\text{sub}}}$. The graph encoder is used to learn graph embedding $Z$ and $Z_{\text{sub}}$ from $G$ or $G_{\text{sub}}$. The mutual information estimator assesses the mutual information between $I(G_{\text{sub}}, G)$ between $G$ and $G_{\text{sub}}$.
• Figure 3: SVM training on the single-cohort dataset and multi-cohort dataset, result in 30% and 35% overlap cross dataset respectively. The overlap of BSNIP and UCLA is focused on Amygdala_R, Calcarine_L, Calcarine_R, Postcentral_L, Postcentral_R, and Caudate_R. The overlap brain regions of multi-cohort training includes Frontal_Inf_Oper_R, Supp_Motor_Area_L, Paracentral_Lobule_R, Caudate_L) and Caudate_R, Calcarine_L, Calcarine_R
• Figure 4: The BrainIB++ model demonstrated 40% overlap in distinct node preferences across both the BSNIP and UCLA datasets during multi-cohort training. Fig.(a) and Fig.(b) illustrate the nodes with the highest probabilities selected by the well-trained BrainIB++ subgraph generator for the BSNIP dataset and the UCLA dataset, respectively. Fig.(c) displays the five common nodes identified in both (a) and (b), including Supp_Motor_Area_L, Calcarine_R, Occipital_Inf_R, Fusiform_R, Postcentral_R. Supp_Motor_Area_L and Calcarine_R appeared more than once within the overlapping nodes.