[Paper Review] Classification of developmental and brain disorders via graph convolutional aggregation

Ibrahim Salim, A. Ben Hamza|arXiv (Cornell University)|Nov 13, 2023

Autism Spectrum Disorder Research4 citations

TL;DR

This paper proposes an aggregator normalization graph convolutional network (AN-GCN) that improves classification of autism spectrum disorder and Alzheimer’s disease by fusing imaging and non-imaging features into a population graph with node and edge features. By integrating skip connections, identity mapping, and graph sampling-based aggregation, the model achieves 50% and 13.56% relative accuracy gains over GCN on ABIDE and ADNI datasets, respectively, with no increase in computational complexity.

ABSTRACT

While graph convolution based methods have become the de-facto standard for graph representation learning, their applications to disease prediction tasks remain quite limited, particularly in the classification of neurodevelopmental and neurodegenerative brain disorders. In this paper, we introduce an aggregator normalization graph convolutional network by leveraging aggregation in graph sampling, as well as skip connections and identity mapping. The proposed model learns discriminative graph node representations by incorporating both imaging and non-imaging features into the graph nodes and edges, respectively, with the aim of augmenting predictive capabilities and providing a holistic perspective on the underlying mechanisms of brain disorders. Skip connections enable the direct flow of information from the input features to later layers of the network, while identity mapping helps maintain the structural information of the graph during feature learning. We benchmark our model against several recent baseline methods on two large datasets, Autism Brain Imaging Data Exchange (ABIDE) and Alzheimer's Disease Neuroimaging Initiative (ADNI), for the prediction of autism spectrum disorder and Alzheimer's disease, respectively. Experimental results demonstrate the competitive performance of our approach in comparison with recent baselines in terms of several evaluation metrics, achieving relative improvements of 50% and 13.56% in classification accuracy over graph convolutional networks on ABIDE and ADNI, respectively.

Motivation & Objective

• To improve disease prediction in neurodevelopmental and neurodegenerative disorders using graph-based representation learning.
• To address over-smoothing in deep graph convolutional networks through architectural innovations.
• To integrate both imaging and non-imaging features into a unified graph structure for holistic brain disorder modeling.
• To maintain computational efficiency while enhancing predictive performance on large-scale neuroimaging datasets.
• To enable interpretable and scalable models for early diagnosis and personalized treatment planning.

Proposed method

• The model constructs a population graph where nodes represent subjects with imaging and non-imaging features, and edges encode pairwise similarities between subjects.
• Graph convolution is performed via a layer-wise aggregation rule that leverages normalized graph sampling to reduce variance in feature propagation.
• Skip connections are introduced to allow direct flow of input features to deeper layers, preserving initial feature information.
• Identity mapping is applied to maintain structural integrity of the graph during feature learning, reducing information degradation.
• Aggregator normalization is used to stabilize training and mitigate bias in mini-batch estimation during graph sampling.
• The model maintains the same memory and time complexity as standard GCNs, enabling efficient scaling.

Experimental results

Research questions

• RQ1 Can graph convolutional networks be effectively adapted for classifying developmental and brain disorders using multimodal data?
• RQ2 How does incorporating skip connections and identity mapping improve performance and mitigate over-smoothing in GCNs for brain disorder prediction?
• RQ3 To what extent does graph sampling with aggregator normalization enhance model stability and generalization on neuroimaging datasets?
• RQ4 Does the proposed model outperform standard GCNs and recent baselines in terms of accuracy and robustness on ABIDE and ADNI?
• RQ5 Can the model maintain high performance without increasing computational cost compared to existing GCN variants?

Key findings

• The AN-GCN model achieves a 50% relative improvement in classification accuracy over standard GCN on the ABIDE dataset for autism spectrum disorder prediction.
• On the ADNI dataset, the model shows a 13.56% relative improvement in accuracy compared to GCN for Alzheimer’s disease classification.
• The model maintains the same memory and time complexity as standard GCNs, enabling efficient deployment without performance trade-offs.
• Batch size analysis shows optimal performance at 1000, with larger sizes leading to reduced generalization and smaller sizes risking non-convergence.
• The model demonstrates robustness across datasets, with consistent performance gains despite domain shifts and complex neurological patterns.
• Limitations include reduced interpretability due to skip connections and identity mapping, and potential performance degradation on out-of-distribution data due to domain shift.