DOI: 10.1161/circ.148.suppl_1.348 ISSN: 0009-7322

Abstract 348: Machine Learning Approaches for Early Outcome Risk Stratification Within 6 Hours of Cardiac Arrest

Qingchu Jin, Richard R Riker, Hunter Williams, Teresa May, David B Seder, Raimond L Winslow
  • Physiology (medical)
  • Cardiology and Cardiovascular Medicine

Background: Formal prognostication should be delayed at least 72 hours after cardiac arrest, but early stratification of risk severity provides important information for medical decision making, research enrollment, and precision medicine.

Hypothesis: EEG-derived indices are valid biomarkers of severity of injury after cardiac arrest and contribute to early identification of risk.

Aims: To assess the relative impact of early clinical data and processed EEG indices in the first 6 hours after recovery of spontaneous circulation (ROSC).

Methods: Data from the first 6 hours after ROSC in the International Cardiac Arrest Registry (INTCAR) from Maine Medical Center including demographics, cardiac arrest data, admission vital signs and laboratory results and processed EEG indices at 6 hours after ROSC (bispectral index or BIS6 and suppression ratio or SR6) were extracted. Cerebral Performance Category (CPC) scores at hospital discharge and 6-month follow-up were dichotomized into good (CPC = 1-2) and poor (CPC = 3-5) outcomes. Three models were trained, one with INTCAR data, a second with EEG data, and the third combining INTCAR and BIS6-SR6. Six machine learning algorithms were applied: Catboost, random forest, Xgboost, Adaboost, support vector machine and logistic regression. Variable importance was analyzed by the decrease of impurity of random forest.

Results: Among 913 included patients, median age was 59 years (95%CI 23-85), 623 (68%) were male, initial rhythm was shockable in 405 (44%) patients, and average time to ROSC was 23 (95%CI 3 - 73) minutes. Survival to hospital discharge occurred in 41% (364/893), and poor outcome occurred in 72.7% at 6-month follow-up (607/835). For CPC at discharge, area-under-the curve (AUC) was 0.85 for the combined INTCAR and EEG data, greater than INTCAR (AUC = 0.80) or EEG (AUC = 0.77). For the CPC at follow-up, the AUC was 0.88 for the combined data, greater than INTCAR (AUC = 0.83) or EEG (AUC = 0.79). Variable ranking analysis showed that BIS6 and SR6 were the two most informative features for both outcomes.

Conclusions: The processed EEG variables BIS6 and SR6 significantly contribute to risk identification in the first 6 hours after ROSC, and combined with early clinical data provide reliable risk stratification.

More from our Archive