Bossone Research Center, Room 709, located at 32nd and Market Streets.
BIOMED PhD Thesis Defense
Machine Learning Enhanced Translational Research Methods Refine Novel Biomarker for Human Epilepsy
Walter Hinds, PhD Candidate, School of Biomedical Engineering, Science and Health Systems, Drexel University
Karen A. Moxon, PhD, Professor, University of California, Davis; and Research Professor, School of Biomedical Engineering, Science and Health Systems, Drexel University
One in three cases of epilepsy do not respond to drug treatment. If left untreated, status epilepticus can have severe symptoms and even lead to death. In some cases, it is possible to identify the location of brain tissue that is generating seizures and surgical removal of this area may result in reduced or even complete stoppage of seizures. Unfortunately, localization of epileptic brain tissue is difficult and complex cases require advanced techniques. Currently, electrocorticography (ECoG) is the ideal tool for identifying epileptic tissue because it has good temporal and spatial resolution. A recently proposed electrophysiological biomarker for epileptogenic tissue is the high frequency oscillation (HFO). In this thesis dissertation, translational research is applied to add greater capabilities for using ECoG to detect HFO’s, primarily as a surgical tool for identifying epileptogenic tissue.
Identifying the epileptogenic zone, i.e., the brain tissue causing seizures, is not straightforward due to multiple reasons. The spatial resolution for ECoG is limited to 2 millimeters by multi-modal co-registration and detecting HFO’s is confounded by frequent false detections (10 - 60%). These issues can affect the localization of epileptogenic zone because HFO biomarkers occur in relatively small numbers and in sparse areas of the brain. Therefore, improving spatial and temporal localizations of the HFO biomarker will aid surgical removal and improve outcomes for cases of intractable epilepsy.
To account for the multi-modal co-registration errors, a novel method was developed using the patient’s post-implant magnetic resonance image (MRI). Although the MRI-MRI intra-modality allows for better registration, typically this scan is not used due to limited visibility of electrode imprints. However, after applying an iterative closest point (ICP) matching algorithm, the electrode artifacts extracted from brain-boundary registered post-implant MRI's can be utilized to enhance the fusion of multi-modal co-registrations. Additionally, this semi-automated method can be applied seamlessly to multiple types of electrodes, i.e., depths, strips, and grids.
To reduce the false detections of HFO’s a machine learning classification algorithm was developed from gold-standard, manually identified samples. First the initial detection procedure was used to find as many HFO candidate events as possible with 97% sensitivity. From this, a pool of ~2,000 hand-labeled events (~1,000 HFO and ~1,000 Noise) was created to train the algorithm and test its performance. The algorithm operates as a multiple classifier system (MCS) created from nine patient’s individually trained, artificial neural network (ANN) classifiers. The MCS was able to classify the events with 89% sensitivity, i.e. preserving almost 9 out of 10 true HFO events, while having the lowest reported false detection rate 5% of all available methods.
The channel-wise false detection rate of seizure onset zones was also reduced 20% by the MCS, while maintaining the same sensitivity. Analysis using patient resections and outcomes found HFO rates detect the epileptogenic zone with a sensitivity of 100% and negative predictive value of 100%. Therefore, the clinical relevance of automatically detected HFO's is promising, especially as a tool for surgical brain mapping. However, extensive analysis of this novel biomarker is still necessary to completely realize its potential diagnostic utility.