EEG Functional Connectivity measures in outcome prediction of postanoxic coma patients
Martin Carrasco Gomez
14 November 2018
Introduction: Early prognostication of neurological outcome in postanoxic comatose patients remains as a challenge, especially after the introduction of therapeutic hypothermia treatment. Functional connectivity stands as one possible contributor to prediction of outcome in this population, given the synaptic failure and neuronal cell death that takes place in hypoxic-ischemic brain injury, as well as its alteration in other disorders of consciousness. In this project, outcome prediction through quantitative analysis of functional connectivity in EEG signals from postanoxic coma patients was performed. In this project, the prognostic value of these measures, as well as its additional value to the state of the art in postanoxic coma outcome prediction, is addressed.
Methods: In a prospective cohort study, the EEGs recorded at 12, 24 and 48 hours after cardiac arrest were analyzed. Outcome was assessed at 6 months and categorized as ”good” (Cerebral Performance Category 1-2) or ”poor” (CPC 3-5). Coherence (COH), corrected imaginary COH, phase locking value (PLV), corrected imaginary PLV, mutual information (MI) were calculated. Elastic nets were used to select the relvant parameters among those, which were then used to predict the outcome of patients through the construction of different machine learning classifier models. Additional machine learning classifiers were trained with a set of EEG general parameters, and a combination of those and functional connectivity features, so to assess the additional value of functional connectivity.
Results: Of the 559 patients included, 46% had a good outcome. A sensitivity of 50.8 ± 4.4% (Mean ± standard deviation) and a specificity of 99.9±0.5 was obtained at 12 hours after resuscitation using a functional connectivitybased model for prediction of poor outcome. The best model based on the EEG general parameters alone (12 & 48 hours data) for poor outcome prediction presented a sensitivity of 32.6±13.9% and a specificity of 100,0±0.0, while the model based in the combination of these parameters and functional connectivity measures showed a sensitivity of 69.3±7.4% and a specificity of 100,0±0.5. The best model for good outcome prediction was based on functional connectivity measures from all available data (12,24 & 48 hours), obtaining a sensitivity of 68.8±0.0 and a specificity of 95.1 ± 0.0.
Conclusions: Functional connectivity stands as a reliable predictor of postanoxic coma by itself. In addition, these measures provided with generalisability to the classification models trained with both EEG general parameters and functional connectivity features. Increased functional connectivity over all frequency bands was observed in poor outcome patients after 24 hours, while an increased theta and alpha connectivity was reported in good outcome patients at 12 hours after resuscitation. Poor outcome patients showed networks that were configured towards small-world architecture, but inefficient, when compared to good outcome patients.