We are two PhD students in biomedical engineering at Universitat Pompeu Fabra in Barcelona, forming the UPFantastic team to compete in the 2023 PhysioNet Challenge with the support of our supervisors and research groups. Our interpretable approach uses electroencephalographic (EEG) signals to classify survival chances in cardiac arrest, prioritizing clinician understanding by identifying relevant features.
Following the challenge objectives, we have conducted an exploratory stage to determine the relevant features for classifying patients as having poor or good outcomes. We applied signal analysis techniques to the EEG, we started by analyzing classical band-energy methods. Specifically, we selected the alpha and beta bands to be the most pronounced during seizure-like periods. We band-pass filtered the recordings in these ranges and quantified the phase-locking degree for signals in pairs and in groups.
Using the extracted features and patient data, we trained algorithms for classification (poor vs. good outcome) and regression (CPC score), achieving an average position on the classification ladder with scores of 0.30, 0.55, 0.43, 0.36 for 12, 24, 48, 72 hours, respectively. In our initial tests, we focused on feature extraction and did not modify the provided classification/regression models, leaving room for improvement in our approach.
As following steps, we aim to enhance the model complexity, by creating a cascade-like architecture using our early hours results. Moreover, we plan to perform trajectory analysis to select the most important features in a reduced dimensionality space (obtained using Multiple Kernel Learning) and regressing their behavior back to the original space, given the time organization of the data.
Our approach has potential for improvement as we prioritized feature extraction over model tuning. It involves multiple stages but is still highly interpretable, in contrast to black-box methods. We are excited to continue working on it and share our findings at the CinC 2023 conference.