Abstract
Predicting neurological recovery in comatose patients following cardiac arrest is a critical clinical challenge. This study investigates the use of Transformer-based neural networks to automate this prediction using Electroencephalogram (EEG) data from the I-CARE Database. We developed four experimental setups (TRS1-TRS4) to evaluate the impact of different temporal windows. Our results demonstrate that a model focusing solely on the most recent hour of EEG data (TRS1) outperforms models utilizing longer, but more incomplete, 24-72 hour windows. This "less is more" phenomenon is attributed to the high prevalence of missing data in early clinical recordings.
Introduction
Cardiac arrest often results in oxygen deprivation to the brain, leading to comatose states. Clinicians must accurately predict whether a patient will achieve a "good" neurological recovery or face a "bad" outcome, such as a permanent vegetative state or death. While EEG provides vital insights into brain activity, its analysis is complicated by its high-dimensional, sequential nature and the frequent presence of data gaps in real-world clinical settings. This research aims to develop a robust automated prediction system capable of handling these constraints.
Methods
The proposed system utilizes a "Vanilla Transformer" encoder-only architecture, leveraging the self-attention mechanism to identify patterns in sequential EEG signals. As part of the 2023 PhysioNet Challenge, two parallel models were trained: one for binary outcome prediction (Good vs. Bad) and another for predicting the Cerebral Performance Category (CPC) score.
Four experimental configurations were tested:
- TRS1: The primary model, utilizing only the most recent hour of EEG data with a simplified feature set (4 channels, PSD across four frequency bands).
- TRS2-TRS4: Models utilizing the first 24, 48, and 72 hours of data, respectively, with a more complex 19-channel feature set.
The models were trained on the I-CARE Database (1,020 patients), using a 60/10/30 split for training, validation, and testing.
Key Insights and Data Quality
A significant finding of this study is the "Less is More" paradox: the TRS1 model, despite having access to the least amount of temporal data, achieved the highest performance. This is explained by the distribution of available data over time.
As shown in the data availability plot, the initial hours of recording are often missing or incomplete. By focusing on the most recent hour, the TRS1 model avoids the noise and bias introduced by imputation methods required for the sparser early-hour data.
Results
The TRS1 model achieved a score of 0.37 on the official PhysioNet Challenge test set, ranking 29th globally. Internal cross-validation revealed an AUC-ROC of 0.77 and an F1-score of 0.79 for TRS1, significantly outperforming the more complex TRS2-TRS4 models. Crucially, the model maintained high sensitivity at a fixed False Positive Rate of 0.05, a vital metric for clinical safety to avoid false "bad" predictions.
Conclusion
This research highlights the importance of data quality over quantity in clinical deep learning. While Transformers are powerful for sequential data, their performance is heavily dependent on the integrity of the input. Future work will focus on more sophisticated imputation techniques and training on larger multi-center datasets to further refine these predictive models.
Summary updated