P.019 A machine learning approach to asymmetric burst suppression and refractory status epilepticus outcome

Abstract

Background: Treatment of refractory status epilepticus (RSE) is often titrated to achieve EEG burst suppression. However, optimal burst suppression characteristics are largely unknown. We used an unsupervised machine learning algorithm to predict RSE outcome based on the quantitative burst suppression ratio (QBSR). Methods: We conducted principal component analysis (PCA) as a linear combination of 22 QBSR features from non-anoxic adult RSE patients at the Winnipeg Health Sciences Centre. We also determined the most predictive components that significantly differed between survivors and non-survivors. Results: Using 135,765 QBSRs from 7 survivors and 10 non-survivors, PCA identified a predominantly non-survivor cluster of 8 patients (75% non-survivors). The first 2 PCA components comprised 75% data variance. The most important first component feature was skewness of QBSR distribution in the right or left hemisphere (0.52 each). The most important second component feature was third QBSR quantile of the left hemisphere (0.49). Only right hemispheric QBSR features significantly differed between groups: QBSR skewness for the first component (Benjamini-Hochberg adjusted p=0.038) and average QBSR for the second component (0.32, Benjamini-Hochberg adjusted p=0.046). Conclusions: Our pilot study shows that RSE patient survival may be impacted by QBSR, with differential hemispheric EEG burst suppression characteristics predicting poor RSE outcome.