Authors :
Presenting Author: Karen Wurzel, BS – Marquette University
Daniel Ackom, MS – Marquette University and Medical College of Wisconsin Joint Department of Biomedical Engineering
Andrew Crow, BA – The Medical College of Wisconsin, Milwaukee
Avantika Singh, MD – Medical College of Wisconsin
Pradeep Javarayee, MD MBA – The Medical College of Wisconsin, Milwaukee
Scott Beardsley, PhD – Marquette University and Medical College of Wisconsin Joint Department of Biomedical Engineering
Richard Povinelli, Ph.D., P.E. – Marquette University — OPUS College of Engineering
Rationale:
Early effective treatment for Infantile Epileptic Spasms Syndrome (IESS) is vital, yet predicting medication response in non-lesional cases remains challenging. Neural networks using MRI-derived structural connectomes have shown promise in distinguishing refractory from responsive epilepsy in adults. These connectomes, modeled as weighted graphs of brain region connectivity, are well suited for Graph Neural Networks (GNNs). Building on our recently published findings that structural connectivity may predict medication response in IESS, we aim to harness the representational power of GNNs to improve prediction accuracy and clinical utility.
Methods:
Structural connectomes are generated using pre-treatment T1-weighted and diffusion MRI data from IESS patients with the UNC/UMN Baby Connectome Project atlas. Figure 1a illustrates the structural connectome pipeline. From each connectome, graph-based metrics—including Katz centrality, betweenness centrality, degree centrality, closeness centrality, and clustering coefficient— are computed and, along with patient age at scan time, used as node features in the GNN model.
To mitigate overfitting given the small sample size, we employed leave-one-out cross-validation. The dataset is imbalanced, with 20 drug-responsive and 6 refractory patients. To address this, we apply the SMOTE technique (k=3) to synthetically augment the minority class during training (Figure 1b). To preserve biological plausibility, edge lists for SMOTE-generated samples are randomly drawn from existing edge lists of the original minority class.
Results:
The study includes a total of 26 patients (17 females) between the ages of 4 and 12 months, with a mean age of 6.8 ± 2.12 months. Eleven patients (42.3%) respond to first-line medication, while nine patients (34.6%) and two patients (7.6%) respond to second- and third-line treatments, respectively. Table 1 summarizes the patients’ clinical profiles and medication response.
The GNN model is evaluated using leave-one-out cross-validation (Figure 1b) across all 26 original patient samples. The GNN model architecture consists of two graph convolutional layers followed by a fully connected layer (Figure 1c).
Performance metrics include accuracy, precision, recall/sensitivity, specificity, F1 score, negative predictive value, and false positive rate with values of 0.85, 0.90, 0.90, 0.67, 0.90, 0.67, and 0.33, respectively. The confusion matrix summarizing the results is shown in Figure 1d.
Conclusions:
GNN models show promise as a computational tool for aiding early differentiation between drug-responsive and refractory IESS patients, potentially supporting more informed clinical decision-making. However, limitations of the study include a small sample size, the imbalance in the size of the refractory and responsive classes, and the assumption of feature independence when generating synthetic refractory training samples. Thus, future work will investigate GraphSMOTE to generate more robust synthetic refractory training samples with the aim of improving the false positive rate.
Funding: Children’s Wisconsin Research Institute Foundation Grant and Advancing a Healthier Wisconsin Endowment Grant.