Automated Machine Learning-based diagnosis of impaired consciousness: cross-center and protocol generalization of EEG biomarkers.

To date, electroencephalography (EEG) and machine learning begin to assist the diagnostics of post-comatose conditions of impaired consciousness after severe brain injury. Recent findings suggest that a variety of EEG-markers are of complementary diagnostic value and can act synergically through multivariate predictive modeling. However, it is unclear to which extent such models depend on specific EEG configurations or protocols, and an assessment of bias and generality through application to independent data is pending. Here we probed the capacity of EEG markers of consciousness and predictive models to discriminate the unconscious from the minimally conscious patients, using different sensor configurations and recording lengths from 141 EEG. Moreover, we demonstrated prospective validity by testing our model on new data (n=108) and observed negligible cross-validation bias. Furthermore, our model could generalize to resting state clinical routine EEG from an independent clinical center (n=48), suggesting that it captures consciousness-specific patterns. These findings imply that EEG signatures of consciousness can be reliably extracted from different contexts and combined into coherent predictive models, encouraging future efforts in large-scale data-driven clinical neuroscience. Finally, these findings are now translated to a web server in which clinicians upload recordings and obtain an automated report with EEG markers and a prediction of the state of consciousness.