Machine Learning-Based Feature Quantification of Clinical High-Frequency Oscillations

Biomedical Engineering Master's Defense - Neha Sara John

WHERE: Lurie Biomedical Engineering (formerly ATL), 1170 show on map

WHEN: December 9, 2022 10:00 am-11:00 amADD TO CALENDAR

Machine Learning-Based Feature Quantification of Clinical High-Frequency Oscillations: Biomedical Engineering Master's Defense - Neha Sara John

Abstract:
The suspense of not knowing when and how a seizure may occur is one of the most exhaustive aspects of an epileptic patient’s life. To address this problem, researchers have been investigating novel biomarkers to gain information about seizure generation and epileptic networks for decades. With modern advancements in recording equipment and computational power, biomarkers that identified from the electrical activity recorded by intracranial electrodes, such as High-Frequency Oscillations (HFOs) have gained traction and are utilized to predict seizure onset zone (SOZ) locations within the brain. However, HFOs are low amplitude waveforms with a time period less than 1 millisecond whose acquisition process can result in noise artifacts. The EEG signals undergo filtering to isolate the HFO events within the 100 – 500 Hz range; a process that can produce false positives due to the occurrence of Gibb’s Phenomenon. Additionally, these filters can also mask the occurrence of artifacts such as head movement and electrical noise. Thus, to address this problem, a machine learning classifier was developed to distinguish events that are clearly artifact from the true HFOs. In this study, logistic regression models were designed to distinguish HFOs naturally occurring in the brain from false positive resulting artifacts amounting to a Positive Predictive Value (PPV) of 84.59%. The correlations of these correctly identified HFOs with inter-patient variability and feature prioritization were also analyzed in a repeated measures study and the significance of each feature was computed. Future directions of this study would follow the estimation of more features from the HFO data to refine the algorithm and improve the precision of identifying true-positive HFO detections. This will heighten the quality of the HFOs being detected in real-time continuous EEG recordings on translation into a clinical environment in the future.

Zoom Link: https://umich.zoom.us/j/96623464149 Meeting ID: 966 2346 4149

Committee Chair(s):
Dr. William Stacey