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

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

TL;DR

Researchers at Michigan State University have developed OTGCN, a novel framework that combines Graph Convolutional Networks (GCNs) with Optimal Transport (OT) to detect Autism Spectrum Disorder (ASD) across different clinical sites. By mathematically "transporting" data distributions from a labeled source site to an unlabeled target site, the model overcomes the "domain drift" caused by different MRI scanners and patient demographics, significantly outperforming existing state-of-the-art methods.

What Problem Does This Paper Address?

Machine learning models for medical diagnosis often suffer from a "site-dependency" problem. A model trained on brain scans from one hospital might fail miserably when applied to scans from another hospital because of differences in imaging equipment (e.g., Tesla strength of MRI machines), procedural protocols, or local patient demographics. This is known as domain drift or concept drift.

In the context of Autism Spectrum Disorder (ASD), researchers often use "population graphs," where nodes represent individuals and edges represent similarities in their brain connectivity or phenotypic data (age, sex, etc.). The challenge is Cross-Network Node Classification (CNNC): how do you accurately classify nodes in a new "target" network (a new hospital) when you have plenty of labeled data from a "source" network but zero labels for the new one?

Methods

The authors introduce OTGCN (Optimal Transport Graph Convolutional Network), which uses a three-pronged approach:

  1. Structural Embedding (GCN): Standard GCN layers capture how a patient relates to their "neighbors" in the population graph.
  2. Node Feature Transformation (NFT): The authors observed that population graphs often have "low homophily"—meaning a patient's neighbors aren't always similar to them. To fix this, they added NFT layers that independently learn non-linear embeddings of a patient's raw features, ensuring individual data isn't "washed out" by noisy graph connections.
  3. Optimal Transport (OT) Layer: This is the core innovation. Using the Sinkhorn algorithm, the model calculates a "transportation plan" to map the distribution of the source site's embeddings onto the target site's distribution. This aligns the two datasets in a shared latent space, allowing a single classifier to work effectively on both.

The training happens in two stages: a pretraining phase to seed the weights using standard cross-entropy loss, followed by an OT training phase that minimizes a joint loss function (classification accuracy + optimal transport cost).

Datasets

The study utilizes the ABIDE (Autism Brain Imaging Data Exchange) dataset, a large-scale collection of resting-state functional MRI (fMRI) and phenotypic data.

  • Source Data: 804 samples pooled from 17 different collection sites.
  • Target Data: 67 samples from 3 specific sites (YALE, CALTECH, and CMU) held out to simulate a new, unlabeled clinical environment.
  • Features: Node features were derived from phenotypic information, while edges were constructed based on the similarity of fMRI image connectivity.

Key Insights and Ideas

  • The "Perils of Drift": The paper visually demonstrates (via t-SNE plots) that a decision boundary perfectly separating "Autism" vs. "Control" in one hospital will likely cut right through the middle of the data in another hospital if you don't account for distribution shifts.
  • GCNs aren't a Silver Bullet: Standard GCNs rely on the "homophily principle" (similar nodes stay together). In medical population graphs, this isn't always true. The addition of the NFT layer is a critical "aha moment," as it allows the model to balance individual patient data with neighborhood context.
  • Barycentric Mapping: By using OT to "transport" source features to the target domain, the model essentially creates a "synthetic" version of the source data that looks like it was collected at the target site, making the final classification task much easier.

Key Results

OTGCN demonstrated a massive performance leap over both standard GCNs and existing cross-network baselines:

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

The results show that while a standard GCN performs barely better than a coin flip (0.502 F1) when faced with site-to-site drift, OTGCN maintains near-perfect accuracy by successfully aligning the data distributions.

Limitations and Future Work

  • Target Sample Size: The target dataset was relatively small (67 samples). Testing on larger, more diverse target networks would further validate the model's robustness.
  • Graph Construction: The model's performance is still somewhat dependent on how the initial population graph is built (e.g., the threshold for similarity).
  • Computational Cost: While the Sinkhorn algorithm makes Optimal Transport feasible, calculating transportation plans for extremely large graphs (tens of thousands of nodes) could become a bottleneck.
  • Future Direction: The authors suggest exploring different types of "cost functions" within the OT layer to see if other distance metrics (beyond squared $l_2$ distance) yield better alignments for medical data.

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