Background: Cardiac arrest is a major medical emergency. Even after successful resuscitation, patients can suffer from severe brain injury and become comatose. It would be important to identify patients who have a chance to recover consciousness. Monitoring electroencephalography (EEG) can be used to help predict prognosis, but the interpretation of the EEG requires expertise. This study aimed to train a machine learning model that predicts the prognosis of comatose patients after cardiac arrest.
Method: We used data from the International Cardiac Arrest REsearch consortium (I-CARE) database. It contains EEG signals from 1020 patients who have been resuscitated after cardiac arrest but remain comatose: Out of these 1020 patients, 607 were published as a training set, and 413 were kept as the hidden validation and test sets for the George B. Moody Physionet Challenge 2023. We calculated expert features from the EEG signals. The calculated features from frequency ranges (α, β, θ, δ waves, and slow wave activity) included demographics, spectral power, spectral entropy, power of the signal, auto correlation, and statistical features from the 18 EEG channels. We trained a Random Forest model which was evaluated using the provided challenge scoring metric.
Results: For predicting the outcome of patients, our group UTU_EEGenies submitted a model which received the following challenge scores of 0.09, 0.12, 0.37, and 0.58 for 12, 24, 48, and 72 hours, respectively. This resulted in the 21st place on the challenge leader board.