Using Embedding Extractor and Transformer Encoder for Predicting Neurological Recovery from Coma After Cardiac Arrest

0582514 - ÚPT 2024 RIV US eng C - Conference Paper (international conference)
Pavlus, Ján - Pijáčková, Kristýna - Koščová, Zuzana - Smíšek, Radovan - Viščor, Ivo - Trávníček, Vojtěch - Nejedlý, Petr - Plešinger, Filip
Using Embedding Extractor and Transformer Encoder for Predicting Neurological Recovery from Coma After Cardiac Arrest.
2023 Computing in Cardiology (CinC). New York: IEEE, 2023. ISBN 979-8-3503-8252-5. ISSN 2325-8861. E-ISSN 2325-887X.
[Computing in Cardiology 2023 /50./. Atlanta (US), 01.10.2023-04.10.2023]
R&D Projects: GA TA ČR(CZ) FW06010766
Institutional support: RVO:68081731
Keywords : Deep learning * Pipelines * Cardiac arrest * Medical services * Predictive models * Electrocardiography * Brain modeling
OECD category: Medical engineering
https://www.cinc.org/archives/2023/pdf/CinC2023-054.pdf https://ieeexplore.ieee.org/document/10364171

This research presents a deep-learning framework designed to forecast neurological recovery following a cardiac arrest-induced coma. The framework is created by the team ISIBrno-AIMT as part of the Predicting Neurological Recovery from Coma After Cardiac Arrest: The George B. Moody PhysioNet Challenge 2023. Our approach involves a two-stage model: initially, the model derives low-dimensional embeddings from short electroencephalogram (EEG) segments (5 minutes), and subsequently, it combines the temporal progression (72 hours) of these embeddings to yield a comprehensive likelihood assessment of recovery outcomes. Regrettably, our submission was not evaluated in the ranking phase due to issues with the Docker pipeline.
Permanent Link: https://hdl.handle.net/11104/0350908