Authors :
Presenting Author: Karthik Gopalakrishnan, – Oregon State University
Shini Renjith, Ph.D. – Oregon State University; Tobias Loddenkemper, MD – Harvard Medical School; Daniel Friedman, MD – 3NYU Langone Medical Center; Mark Spitz, MD – University of Colorado Health; Mitchell Frankel, Ph.D. – Epitel,Inc; Mark Lehmkuhle, Ph.D. – Epitel,Inc; V John Mathews, Ph.D. – Oregon State University
Rationale:
Machine learning (ML) algorithms have found varying levels of success in detecting electrographic seizures using electroencephalogram (EEG). However, these algorithms are not suitable for continuous monitoring of patients during activities of daily life as they do not generalize well across patients and their use of bulky wired sensors makes them cumbersome for continuous, ambulatory use. This work presents a novel ML framework that uses scalp EEG recorded from a single wearable EEG sensor to detect common types of electrographic seizures and is capable of generalizing across patients.
Methods:
We developed pipelines to separately detect three broad classes of seizures: tonic-clonic seizures (TC); generalized absence seizures (GA); and focal seizures with impaired awareness (FIA). Scalp EEG was collected from 251 patients who were simultaneously monitored by a standard-of-care 19+ wired-channel video-EEG monitoring system and the single-channel wireless sensor. Out of these, 33 patients had TC seizures, 12 had GA seizures, and 36 had FIA seizures. Known seizure event start and stop times were determined during an epileptologist standard-of-care video-EEG review. The general pipeline for detecting electrographic seizures is shown in Figure 1. The preprocessed EEGs were standardized using an adaptive estimate of the standard deviation of the noise in the data, and the features were normalized using an estimate of the auto-correlation of the features in the training set. The system applied morphological filters to the estimated labels to create a set of discrete detected events. The feature set, the classifier, and the postprocessing parameters were tailored separately for each pipeline.
Results:
The TC pipeline was trained on data recorded from 26 patients and evaluated on data from 7 patients. Similarly, the GA pipeline was trained on data from 10 patients, and evaluated on two patients, while the FIA pipeline was trained on 29 patients and evaluated on seven patients. Each of the evaluation datasets also contained 10 patients who showed no notable electrographic seizure activity. Our framework was able to achieve sensitivities of 100%, 89%, and 68%, and false detection rates (FDR) of 0.13/hr, 1.01/hr, and 0.68/hr for TC, GA, and FIA seizures, respectively.
Conclusions:
This work presented an ML framework to detect electrographic seizures using a single wearable EEG sensor. New data standardization and feature normalization schemes and a morphological filter-based post-processing system were developed for this framework. Experimental results showed that our method detected TC and GA seizures with relatively high sensitivity and low FDR. FIA seizures were more difficult to detect. Such seizures result in different clinical and electrographic manifestations across individuals and therefore may require much larger training sets to achieve more accurate results, especially in low-channel-count situations.
Funding:
This work is supported by NIH NINDS grant U44NS121562.