Population Graph Cross-Network Node Classification for Autism Detection Across Sample Groups

Anna Stephens, Francisco Santos, Pang-Ning Tan, Abdol-Hossein Esfahanian

Abstract

Machine learning models for medical diagnosis, such as Autism Spectrum Disorder (ASD) detection, often struggle with "site-dependency" where models trained on one clinical site fail on another due to domain drift. We introduce OTGCN (Optimal Transport Graph Convolutional Network), a novel framework that combines Graph Convolutional Networks (GCNs) with Optimal Transport (OT) to align data distributions across different sites. By leveraging a three-pronged approach—structural embedding, node feature transformation, and an OT alignment layer—OTGCN achieves state-of-the-art performance in cross-network node classification. Evaluations on the ABIDE dataset demonstrate that OTGCN significantly outperforms existing baselines, maintaining high accuracy even when target labels are unavailable.

Introduction

The application of machine learning to medical imaging is frequently hindered by domain drift (or concept drift). In ASD research, differences in MRI scanner strength, procedural protocols, and patient demographics across hospitals create distribution shifts that degrade model generalizability. This problem is particularly acute in Cross-Network Node Classification (CNNC), where a model must classify individuals in a new, unlabeled "target" network using knowledge from a labeled "source" network.

Standard population graphs represent individuals as nodes and their similarities as edges. However, simple graph-based methods often fail to account for the inherent variability between clinical sites, leading to misaligned decision boundaries. The OTGCN framework was developed to bridge this gap by mathematically "transporting" data distributions from a labeled source site to an unlabeled target site.

Methods

The OTGCN framework addresses domain drift through a multi-stage architecture designed to balance local patient features with global graph structure.

Architecture Components

  1. Structural Embedding (GCN): Standard GCN layers are employed to capture the relational context of patients within the population graph.
  2. Node Feature Transformation (NFT): Recognizing that population graphs often exhibit "low homophily"—where neighbors are not necessarily similar—the authors integrated NFT layers. These layers learn non-linear embeddings of raw patient features independently, preventing individual data from being obscured by noisy graph connections.
  3. Optimal Transport (OT) Layer: This core component utilizes the Sinkhorn algorithm to compute a transportation plan that maps the source site's embedding distribution onto the target site's distribution. This alignment allows a single classifier to generalize across both domains.

Training Procedure

The model is trained in two distinct phases: a pretraining phase to initialize weights using standard cross-entropy loss on source data, followed by a full training phase that minimizes a joint loss function comprising classification accuracy and the optimal transport cost.

The OTGCN architecture featuring pretraining and full training phases with Optimal Transport integration.

Datasets

The study utilized the ABIDE (Autism Brain Imaging Data Exchange) dataset, which includes resting-state functional MRI (fMRI) and phenotypic data.

  • Source Data: 804 samples pooled from 17 different collection sites.
  • Target Data: 67 samples from three specific sites (YALE, CALTECH, and CMU) held out to simulate an unlabeled clinical environment.
  • Features: Node features were derived from phenotypic data, while edges were constructed based on fMRI connectivity similarity.

Results and Discussion

OTGCN demonstrated a substantial performance advantage over both standard GCNs and specialized cross-network baselines. While a standard GCN performed near chance levels (0.502 F1) due to site-to-site drift, OTGCN maintained near-perfect accuracy by successfully aligning the data distributions.

Method Macro F1 Score Micro F1 Score
OTGCN (Ours) 0.979 0.979
ACDNE 0.943 0.943
GCN 0.502 0.558
AdaGCN 0.278 0.371

Visual analysis via t-SNE plots confirms the efficacy of the OT layer. Before transport, the source and target distributions are disjoint, making a common decision boundary ineffective. After transport, the distributions are aligned in a shared latent space, facilitating accurate classification of the target nodes.

t-SNE visualization of feature distributions before and after Optimal Transport alignment.

Conclusion

OTGCN provides a robust solution for cross-site ASD detection by integrating Optimal Transport into a graph-based learning framework. By successfully mitigating domain drift, the model enables the application of diagnostic tools to new clinical environments without requiring local labeled data. Future work will focus on scaling the framework to larger graphs and exploring alternative OT cost functions to further refine alignment in complex medical datasets.

Summary updated

BibTeX
@inproceedings{stephens2023population,
  author = {Anna Stephens and Francisco Santos and Pang-Ning Tan and Abdol-Hossein Esfahanian},
  title = {{Population Graph Cross-Network Node Classification for Autism Detection Across Sample Groups}},
  booktitle = {Workshops @ ICDM},
  year = {2023},
  doi = {10.1109/icdmw60847.2023.00050},
}
Figure