Abstracts

Rationale: Prognostication for patients with disorders of consciousness (DOCs) is challenging due to their diverse etiologies and underlying pathophysiologies. Advances in the field of machine learning (ML) and the increasing availability of long-term EEG and multimodal data may offer improved methods for monitoring and prognostication. We here propose building a comprehensive latent space EEG map trained on large-scale EEGs across the health-disease continuum to map longitudinal health trajectories for individualized prognostication after cardiac arrest.


Methods:
We develop and train a deep autoencoder-like architecture on large EEG datasets, including sleep EEGs (wake, sleep stages N1-N3, REM) and EEGs across the ictal-interictal continuum to group EEGs across the health-disease continuum in a 128-dimensional latent space. The distances between the groups in this universal map of EEG (UM-EEG) represent a measure of the similarity between EEGs. Next, we project out-of-sample multi-day EEGs from patients with DOCs into UM-EEG to measure the distances to the various clusters over time. We use these latent-space trajectories for prognostication of outcome after cardiac arrest.




Results:
UM-EEG effectively clusters EEG states, including ‘healthy’ (all wake and sleep stages) and numerous 'abnormal' states (ictal and interictal data) in a sematic latent space. Semantic clustering is evidenced by the distinct and meaningful spatial arrangement of clusters capturing natural state transitions and good classification performance of out-of-sample data. Projection of unseen, time-continuous trajectories of patient data with DOCs allows to track the distances to all clusters over time to predict outcome (AUC 0.83, logistic regression). Finally, we analysed which UM-EEG factors were most useful for outcome prediction in DOCs which revealed the fraction of time spent in wake and the average distance to the healthy cluster (wake + all sleep stages) to be most informative, achieving an AUC of 0.82 alone. This result supports the intuitive notion that EEGs exhibiting substantial wake-like activity and minimal deviation from healthy patterns are strong predictors of patient outcomes in comatose cardiac arrest patients.




Conclusions: A universal map of EEG (UM-EEG) provides meaningful representation of EEG across the health-disease continuum. Continuously monitoring of brain states in this latent space map may afford more objective, individualized, automated decision support and monitoring of patients with DOCs and other neuro-intensive care setting over extended periods of time. For DOCs, the distance to ‘healthy’ and time spent in wake-like states contain the most information to predict a patient’s outcome.


Funding: Charité Berlin