Eye Movement Detection Using Electrooculography and Machine Learning in Cardiac Arrest Patients
Chelsea Sykora1, Cameron Hill1, Michael Cronin1, Min Shin2, Stephen Schmugge2, Samuel Tate2, Joseph Sisto1, Charlene Ong1
1Boston University School of Medicine, 2UNC Charlotte
Objective:

To train a machine learning algorithm to identify eye movements from electrooculography (EOG) in a cohort of cardiac arrest (CA) patients.

Background:

Neuroprognostication of comatose post-CA patients is challenging and requires new non-invasive biomarkers to better guide families and limit inappropriate withdrawal of life-sustaining therapies. A promising marker of arousal recovery may be eye movements, since pathways for eye movements and arousal share common anatomic structures. Continuous quantification of eye movements is feasible through electroencephalogram (EEG) with EOG. However, manual quantification is difficult and resource intensive. Our aim was to develop an automated machine learning algorithm to quantify eye movements from EOG. Automated quantification of eye movements would enable further studies of association between eye movements and clinical outcomes.

Design/Methods:

We conducted a retrospective, single-center cohort study of patients over age 18 who underwent standard-of-care EEG/EOG monitoring in the intensive care unit post-CA between 2020-2023. Trained team members manually identified eye movements from approximately 60 minutes of EOG. We trained a long short-term memory (LSTM) algorithm to detect eye movement occurrence on 75% of our cohort. We assessed performance on a separate test set comprising 25% of our cohort using area under the curve (AUC), G-mean, sensitivity, and specificity.

Results:

From a preliminary training set of 31 and test set of 15 patients, we found our algorithm had an AUC of 0.882, G-mean of 0.822, sensitivity of 0.762, and specificity of 0.887. Specificity was diminished in patients with reported myoclonic or epileptiform changes on EEG. Comprehensive results and updated analysis will be presented at the meeting.

Conclusions:

An LSTM algorithm appears to be a highly specific model to detect eye movements in CA patients. Non-invasive automated eye movement detection has the potential to aid in the quantification of eye movements in post-CA patients and evaluate for associations with measures of recovery potential.

10.1212/WNL.0000000000204740