Multi-modal Multi-kernel Graph Learning for Autism Prediction and Biomarker Discovery

TL;DR

Multi-modal data pose challenges for disease prediction due to inter-modal interference and heterogeneous information. The authors propose MMKGL, a two-stage framework with Multi-modal Graph Embedding (MMGE) to build per-modality graphs guided by supervision and function graphs, and Multi-Kernel Graph Learning (MKGL) to extract heterogeneous information via multiple Chebyshev graph convolutions, CKDT fusion, and RAM refinement. A gradient-based analysis identifies discriminative brain regions and subnetworks, providing biologically meaningful biomarkers, and ABIDE experiments show state-of-the-art autism prediction performance. The work highlights the practical impact of adaptive graph fusion and cross-kernel information aggregation for both accurate diagnosis and biomarker discovery in autism.

Abstract

Due to its complexity, graph learning-based multi-modal integration and classification is one of the most challenging obstacles for disease prediction. To effectively offset the negative impact between modalities in the process of multi-modal integration and extract heterogeneous information from graphs, we propose a novel method called MMKGL (Multi-modal Multi-Kernel Graph Learning). For the problem of negative impact between modalities, we propose a multi-modal graph embedding module to construct a multi-modal graph. Different from conventional methods that manually construct static graphs for all modalities, each modality generates a separate graph by adaptive learning, where a function graph and a supervision graph are introduced for optimization during the multi-graph fusion embedding process. We then propose a multi-kernel graph learning module to extract heterogeneous information from the multi-modal graph. The information in the multi-modal graph at different levels is aggregated by convolutional kernels with different receptive field sizes, followed by generating a cross-kernel discovery tensor for disease prediction. Our method is evaluated on the benchmark Autism Brain Imaging Data Exchange (ABIDE) dataset and outperforms the state-of-the-art methods. In addition, discriminative brain regions associated with autism are identified by our model, providing guidance for the study of autism pathology.

Figures (7)

• Figure 1: Overview of the architecture of our framework. Multi-modal graph embedding: Multiple modal graphs $\mathcal{G}_{fc}$, $\mathcal{G}_{anat}$, $\mathcal{G}_{func}$ are generated by adaptive graph construction and then are fused into a Multi-modal graph $\mathcal{G}$. The function graph $S_{FG}$ and supervision graph $S_{SG}$ are introduced in the fusion process for optimization. Multi-kernel graph learning: The heterogeneous information of the Multi-modal graph $\mathcal{G}$ is extracted using a graph convolutional network with different convolutional kernel sizes and fused by generating a cross-kernel discovery tensor T for predicting autism spectrum disordor (ASD) and typical control (TC). In addition, relationship attention mechanism (RAM) is used to adjust the weights of the adjacency matrix A.
• Figure 2: Performance comparison of feature fusion method in multi-kernel graph learning and traditional methods.
• Figure 3: (a) Performance comparison of single convolutional kernel receptive field size. (b) Performance comparison of multiple combinations of convolutional kernels with different receptive field sizes.
• Figure 4: (a) Accuracy comparison of multiple methods with different training set ratio. (b) Performance Comparison with different feature input dimensions of MMKGL.
• Figure 5: The figure shows the most discriminative functional connections extracted from the model weights by the gradient-based approach. The thickness of the connections represent their weights. (a) Top 10 discriminative brain functional connectivities. (b) Top 20 discriminative brain functional connectivities.