Machine Learning-informed Subtyping of Generalized Periodic Discharges in Cardiac Arrest (S23.006)

Objective: Identify generalized periodic discharge (GPD) subtypes associated with potential for neurological recovery after cardiac arrest.

Background: GPDs are near invariably associated with poor outcome after cardiac arrest, but a small number of patients may survive with good outcomes.

Design/Methods: We screened EEG data from a cohort of 1,020 comatose patients with cardiac arrest using SParCNet, a neural network for ictal interictal continuum classification trained on EEG data from a general ICU population. Hours of EEG data classified as GPD for at least 15 minutes in the first 120h post-cardiac arrest were included for further analysis. We used 5-fold cross validation and a gradient boosted classifier for good outcome prediction with 27 quantitative EEG features (i.e., spike rate, background continuity index [BCI], and spectral features averaged hourly) acquired from segments with GPD. Good outcome was defined as a Cerebral Performance Category score of 1-2 at 6-months.

Results: 5,525 hours of EEG containing GPDs with an estimated 20 million spikes were analyzed for 300 patients (248 poor and 52 good outcomes). Spike rate in GPD segments was lower in patients with good (median: 0.5 Hz, IQR: 0.61 Hz) vs. poor (0.83 Hz, 1.08 Hz) outcomes. The mean AUC for good outcome prediction using GPD information was 0.8 (+/- 0.11), accuracy of 0.9 (+/-0.03), and specificity of 0.99 (+/-0.01). Following univariate analysis of the most important features in cross-validation, thirty-one percent (N=21/68) of patients with GPDs with BCI > 0.99 and beta to total bandpower ratio < 0.02 were found to have good outcomes.

Conclusions: GPDs with higher continuity and lower beta to total bandpower ratios were associated with a good outcome in cardiac arrest patients. Further subtyping of GPDs associated with potential for neurological recovery may guide decisions about withdrawal of life-sustaining therapies.