Abstracts

Rationale: Critically ill patients with COVID-19 (C-19) infection experience a variety of neurological symptoms including depressed mental status or encephalopathy. Though the etiology of encephalopathy is unknown, it is speculated that activity within, and functional connectivity between, distributed brain networks underlying cognition, attention, and arousal are disrupted. We hypothesized that short and long-range synchronization of neural activity in the brain, measured by phase-locking value (PLV) between electrode pairs in scalp electroencephalography (EEG), is altered in C19-positive (+) individuals compared to a cohort of C19-negative (-) individuals in the same clinical environment and matched for disease severity. To test our hypothesis, we trained a machine learning algorithm on PLV data, arguing that accurate binary classification of C19+ vs. C19- indicates different neural synchrony.




Methods: The study retrospectively analyzes EEG data from ethnically diverse ICU patients in Ohio (OH, Cleveland Clinic) and New York (NY, Kings County Hospital) during wave 1 of the C-19 pandemic (n=58 C19+, n=85 C19-).

EEG evaluations were conducted by trained experts, removing extra physiological artifacts. Data interpretation followed conventional EEG methods. EEG data were transformed into a 36 x 1 matrix to reduce multicollinearity. K-nearest neighbors (KNN) models were trained using cross-validation to optimize hyperparameters.


Results: In the NY dataset, model accuracy reached 74.5% (AUC-ROC 0.74; precision 69.2%; recall 70.2%; specificity 76.7%). Features, which are specific characteristics used by machine learning algorithms to make predictions, consisted of phase locking values (PLVs) indicative of inter- and intra-hemispheric connectivity. Notably, connections such as T6-P4, T3-T4, and F3-T3 were identified, spanning both hemispheres. Similarly, in the OH dataset, model accuracy reached 75.1% (AUC-ROC 0.72; precision 74.7%; recall 70.9%; specificity 78.6%). Here, PLVs predominantly revealed inter- and intra-hemispheric connections, particularly with a prevalence in the left hemisphere, as evidenced by electrode pairs C3-P3, T3-P3, O1-O2, and F7-F8. These findings highlight a commonality between the NY and OH datasets in terms of intra- and inter-hemispheric connectivity patterns, as well as specific electrode locations implicated in distinguishing COVID-positive from COVID-negative cases.

Conclusions: Machine learning models achieved high performance metrics in the binary classification of COVID-19 positive (C19+) versus COVID-19 negative (C19-) patients across two independent major healthcare systems using PLV data computed from EEG. Our results suggest that synchronization of neural activity across distributed networks is a hallmark of depressed mental status in critically ill COVID-19 patients. Although PLV features may differ between populations, asynchronous brain activity, particularly in distinct electrode pairs indicative of inter- and intra-hemispheric connectivity, is a common finding across diverse populations and disease severities.

Funding: No funding support