Abstract: Clinicians primarily rely on the electroencephalogram (EEG) to diagnose subtle and non-convulsive seizures, but EEG interpretation is laborious, and can delay treatment for hours. We demonstrate a method for automating seizure detection. We first transform the EEG waveforms into spectrographic images (n=282), then extract salient descriptors via a computer vision technique, and finally train support vector machines (SVM) on these descriptors with 10-fold cross validation. The SVM achieved 98% accuracy (AUC of ROC curve) in seizure detection

Learning Objective 1: To understand that EEGs can be transformed into spectrographic images, and are then amenable to computer vision and machine learning classification techniques.

Learning Objective 2: To recognize that spectrographic images, and images in general, can be quantitatively and robustly described in higher dimensions using feature extraction techniques.

Learning Objective 3: To recognize that support vector machines are a useful statistical learning method that can accurately classify images based on their descriptors.

Learning Objective 4: To apprepriate the need for automated EEG seizure detection at the bedside.


Peter Yan (Presenter)
Weill Cornell Medical College

Fei Wang, Weill Cornell Medical College
Ramin Zabih, Cornell Tech
Zachary Grinspan, Weill Cornell Medical College

Presentation Materials: