This study aimed to evaluate a computationally efficient machine learning approach using electroencephalography (EEG) on prognosis of comatose cardiac arrest (CA) survivors. Specifically, the study aimed to assess the performance of a robust feature set based on a well-described neurophysiological phenomenon, primarily for early prognosis and secondarily, participating in the PhysioNet Challenge 2023 as team Cerenion, for the scoring times at 12, 24, 48, and 72 hours after the return of spontaneous circulation.
Metadata and EEG data of subset of 608 out of 1020 comatose patients suffering from a recent CA were used to develop the algorithm and to measure its performance. As the crux of the study, feature vectors comprising channel-by-channel root mean square (RMS) power of slow-wave activity (SWA), i.e. frequencies under 1 Hz, in conjunction with time elapsed since CA, were extracted from the patients' EEG recordings. With a random 80/20 split on training and testing data, a random forest classifier was trained and then tested for its performance in predicting the outcome of the patients. The test results were evaluated using Area Under the Receiving Operating Characteristics Curve (AUROC), Area Under the Precision-Recall Curve (AUPRC), accuracy, and F-measure. For the challenge, the results were also scored with a metric weighing false alarms negatively.
Preliminary results for predicting the outcome using the available subset of data were: AUROC 86 %, AUPRC 92 %, accuracy 76 %, F-measure 74 %. Testing on the hidden validation set the corresponding challenge entry scored 0.18, 0.45, 0.52, and 0.60 at the aforementioned times, ranking 18th out of 159 entries.
Preliminary analysis shows promising results in predicting the outcome of a comatose patient after CA using SWA power features in EEG measured of the patient.