Date of Award
Spring 3-19-2026
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Computer Science
First Advisor
Richard Povinelli
Second Advisor
George Corliss
Third Advisor
Scott Beardsley
Abstract
Background: With an incidence rate of approximately three out of 10,000 live births, Infantile Epileptic Spasm Syndrome (IESS) is a rare form of epilepsy with about 2,000 to 2,500 new cases in the United States annually. Each year, thousands of infants and families face uncertainty, often requiring multiple medication regimens before finding an effective treatment. This thesis uses pretreatment structural brain connectivity data to train Graph Neural Networks (GNN) to predict antiepileptic drug response in infants with IESS, aiming to reduce the trial-and-error approach and improve care for those with IESS. Method: The dataset used in this thesis contains de-identified structural connectomes from 26 infants diagnosed with IESS. Node-level features derived from graph-theoretic algorithms are computed for each connectome and used as inputs to the machine learning models. First, a GNN model is trained to classify IESS patients as refractory or responsive to antiepileptic medications. Second, two GNN models are trained to predict response to the first-line treatments, corticosteroids and vigabatrin. Third, ConnectomeSMOTE, a variant of the Synthetic Minority Over-sampling Technique (SMOTE) designed for weighted adjacency matrices, is developed and evaluated to generate synthetic structural connectomes using linear algebra operations to preserve network topology. The performance of ConnectomeSMOTE is compared with the original SMOTE algorithm and a baseline model trained on the imbalanced dataset without oversampling. Oversampling methods are evaluated on a binary IESS diagnostic classification task using the combined dataset of 26 IESS patients and 99 control participants, yielding a total sample of 125 connectomes. Results: The Refractory or Responsive Model achieves an F1 score of 0.90. The Corticosteroid and Vigabatrin Models attain F1 scores of 0.80 and 0.83, and negative predictive values of 0.80 and 0.83, respectively. All oversampling approaches demonstrated discriminatory performance, as measured by AUC-ROC. The Friedman test showed G-mean differences approached statistical significance (p = 0.05), while AUC-ROC, F1-score, and accuracy differences across methods were not significant. Implications: Achieving earlier seizure remission in IESS patients could lead to better cognitive outcomes and improved quality of life. The results highlight the value of GNN models in predicting responses to first-line antiepileptic medications in IESS, supporting more targeted treatment strategies. ConnectomeSMOTE also shows strong potential, but further refinement is needed to maximize its effectiveness.