Predicting Neurological Outcomes of Comatose Cardiac Arrest Patients Using Transformer Neural Networks with EEG Data
Journal
Computing in Cardiology
ISSN
2325-8861
ISBN
[9798350382525]
Date Issued
2023-01-01
Author(s)
Abstract
As part of the 2023 PhysioNet Challenge, our team FINDING_MEMO utilized Transformer to predict outcomes using patient EEG data since it excels at dealing with sequential data like EEG. We mainly used the Transformer encoder block's multi-head self-attention to generate representations from the input and leverage several hidden layers to form the final prediction. Using the latest EEG from every patient, our team achieved the challenge score of 0.42 with the hidden validation set (ranked 36th out of 73 invited teams) and obtained a result of 0.37 (ranked 29th out of 36 qualified teams). Our results show a consistent performance across varying EEG recording durations in both the validation and test set. Our team also had the second-best score when evaluated, with only 12 hours of available recordings in the test set. Such promising results showcase the models' generalizability and clinical potential in predicting outcomes for comatose patients, especially for limited available EEG recordings. © 2023 CinC.
Publisher
Computing in Cardiology
Type
conference paper