Authors: Ryan Dinh Nguyen; Ryan Nguyen, BS; Matthew Smyth, MD; Liang Zhu, PhD; Gretchen Von Allmen, MD; Michael Watkins, MD; Ludovic Pao, BS; Rajan Patel, MD; Jeremy Lankford, MD; Michael Funke, MD, PhD; Manish Shah, MD (Houston, TX)
Introduction: Pediatric epilepsy affects 0.5-1% of children. 30% of children are refractory to medical anticonvulsant therapy. Interictal epileptiform discharge (IED) found in EEG is a common test for epilepsy. IED, however, is specific (78-98%) but not sensitive (25-56%) for epilepsy. This study aims to analyze convolutional neural network (CNN) classification of pediatric refractory epilepsy using resting-state functional magnetic resonance imaging (rsMRI) data. Methods: With IRB approval, the rsMRI and anatomical MRI of 63 refractory epilepsy patients, from Washington University and Memorial Hermann Hospital, and 259 healthy control (HC) patients, from the ADHD-200 data set, were collected. Images were transformed to pediatric atlases in Talairach space, and latency maps of the temporal difference between rsMRI and the global mean signal were calculated using voxel-wise cross-covariance. HC voxel-wise mean and standard deviation latency maps were created and were used to calculate latency z-score maps for HC and epilepsy patients. Latency z-score maps were stratified and randomly split into train, validation and test data sets. A CNN model of two convolutional and a fully connected layer was constructed and trained in Tensorflow. CNN hyper-parameters were optimized via grid-search for best validation accuracy and minimized overfitting, determined by difference between train-validation accuracy. Afterward, the model was run using optimized hyper-parameters and test data. Area under the receiver-operating characteristics curve (AUC) was calculated to evaluate the model’s ability to classify epilepsy in the test set. Results: Epilepsy disease state (epilepsy or no epilepsy) was correctly classified in 80% of test patients with an AUC of 0.75 using an optimized CNN model. Model overfitting was minimized but still present. Conclusion: CNN could classify pediatric epilepsy using rsMRI latency imaging. With further development in model architecture and larger data sets, we hope CNN rsMRI screening could be an adjunct test to diagnose and prognosticate pediatric epilepsy.