Between Patterns and Predictions: Interpretable Latent EEG Representations for Clinical Insights

The article presents a method called Universal Map of EEG (UM-EEG) that preserves clinically meaningful information from electroencephalograms without requiring retraining on new data. The method was tested on three independent patient groups: 576 cardiac arrest patients, 100 subarachnoid hemorrhage patients, and 141 patients with routine clinical EEG recordings. EEG signals were projected into a 128-dimensional space with reference brain states including wakefulness, sleep, seizure activity, and burst suppression. Results showed that patients with favorable outcomes had EEG patterns closer to healthy reference states, while patients with poor outcomes were located in pathological regions. The method achieved ROC-AUC accuracy of 0.83 for predicting cardiac arrest outcomes and 0.76 for subarachnoid hemorrhage. This unified EEG representation enables better interpretation of brain activity and patient outcome prediction.