Introduction: Accurate coma prognostication after cardiac arrest is crucial for patient care and decision-making. In this study, we aim to improve the prediction of neurological recovery for comatose cardiac arrest patients using longitudinal EEG recordings and basic clinical information.
Objectives: During the unofficial phase, we implemented a baseline model using a Random Forest Classifier, which resulted in a score of 0.24. We also experimented with a deep learning approach; however, it produced a false positive (FP) rate higher than 0.05, failing to achieve a valid score. These issues can be attributed to the model-agnostic learning approach which relies purely on data and discards domain knowledge. Therefore, in the official phase of the George B. Moody PhysioNet Challenge 2023, we plan to refine our predictive models by embedding scientific and experiential knowledge into the learning approach.
Methodology: Knowledge augmentation can be implemented at various levels. In this work, we consider training-level integration by incorporating the challenge objectives into the loss function. To this end, we plan to add a penalty term to the loss function that takes into account the violation of the FP<0.05 condition. Additionally, we will perform statistical analysis on the challenge dataset to identify probabilistic dependencies between input variables (such as gender and age) and output variables (such as CPC). This study can reveal important statistical relationships and help us to extract relevant features for our models.
Results: Our primary objective is to formulate the deep learning problem that aligns with the Challenge task as effectively as possible. To achieve this, we plan to incorporate domain knowledge into our models, which we anticipate will lead to an enhanced predictive performance.