Authors :
Presenting Author: Tereza Pridalova, MSc – Mayo Clinic
Filip Mivalt, MSc – Department of Neurology – Mayo Clinic; Vladimir Sladky, MSc – Department of Neurology – Mayo Clinic; Petr Nejedly, MSc – Institute of Scientific Instruments – Czech Academy of Science; Kamila Lepkova, MSc – Faculty of Biomedical Engineering – Czech Technical University in Prague; Petr Klimes, Ing., PhD – Institute of Scientific Instruments – Czech Academy of Science; Jan Cimbalnik, Ing., PhD – International Clinical Research Center, St. Anne’s University Hospital; Benjamin Brinkmann, PhD – Department of Neurology – Mayo Clinic; Vaclav Kremen, MS, PhD – Department of Neurology – Mayo Clinic; Gregory Worrell, MD, PhD – Department of Neurology – Mayo Clinic
Rationale:
Invasive intracranial EEG (iEEG) plays a crucial role in diagnosis and treatment of epilepsy. The increasing volume of recorded data in epilepsy monitoring units (EMUs) with increasing numbers of iEEG channels and recording duration poses a challenge for manual review by epileptologists. While machine and deep learning models excel in specific labeling and classification tasks, their integration into epilepsy monitoring necessitates further investigation.
In response to this challenge, we have developed a system that combines deep and machine learning algorithms with the goal of assisting human experts. This system effectively filters out noise, identifies segments containing relevant pathology, and provides a summary for each iEEG channel. By enabling epileptologists to focus their expertise on crucial data, our system has the potential to enhance the efficiency of interpreting iEEG recordings in EMUs.
Methods:
We developed a framework for visualization and biomarker tracking in EMU iEEG recordings. This framework incorporates a set of deep learning and machine learning approaches to detect noise, pathological iEEG activity (e.g. interictal epileptiform spikes (IESs), high frequency oscillations, and seizures). The methods employed include Conv-GRU models, autoencoders, and a previously validated interictal epileptiform spike detector. Deep learning models were trained using a publicly available iEEG dataset from two institutes, comprising 151,182 three-second segments with expert labeling into four categories (noise: 40,302; pathology: 14,227; physiology: 55,730; powerline interference: 40,922). Results:
We successfully developed an analytical and visualization platform for EMU iEEG recordings, intended for deployment in a prospective seizure-onset-zone identification study. The platform was employed to analyze retrospective recordings from 15 patients with drug-resistant epilepsy, comprising a total of 59 recording days and 839 recording channels. Notably, the detected pathological iEEG activity and IESs exhibited a significantly higher occurrence (p < 0.05