Abstracts

Rationale: The depth of burst suppression EEG patterns is widely held to reflect the brain's global metabolic status, and recent investigations have clarified the underlying neurophysiologic mechanisms. The need to quantify burst suppression frequently arises during continuous EEG (cEEG) monitoring in neuro-ICU patients, e.g. while monitoring depth of anesthesia in pharmacologically induced coma for refractory status epilepticus, and in coma prognostication after hypoxic ischemic injury. Manual cEEG analysis is labor intensive and subjective, necessitating automated methods. We present real-time algorithms for cEEG segmentation and inference of cerebral metabolic state, and validate these across a wide spectrum of burst suppression patterns encountered in ICU cEEG. Methods: A real-time method for segmenting cEEG into burst and suppression epochs is presented based on adaptive variance thresholding. This algorithm is validated against manual segmentations of 20 cEEG recordings by two experienced human electroencephalographers, and robustness of the results to variations in algorithm parameters settings is analyzed. We present an inference algorithm, based on a mathematical model of the mechanisms which generate burst suppression, allowing real-time estimation of the global brain metabolic state. Finally, we present a principled method for fitting this model to burst suppression data from ICU patients undergoing cEEG monitoring. Results: Automated segmentation of burst suppression cEEG records in all 20 cases produced agreement with each human expert that was as good or better than inter-expert agreement. These results were robust to variations in algorithm parameter settings and there was no evidence of ‘overfitting' to the clinical training data. Furthermore, the method was superior to ‘naïve' thresholding methods sometimes advocated in the literature. Moreover, real-time model-based inference of underlying cerebral metabolic state is shown to be feasible across a wide spectrum of ICU EEG burst suppression patterns, and to faithfully reflect key nonlinear EEG features encountered in pharmacologically-induced burst suppression, namely, the nonlinear relation between brain-compartment anesthetic concentration and duration of suppressions in burst suppression EEG. Conclusions: Automated real-time segmentation and inference of cerebral metabolic state from burst suppression cEEG patterns is feasible across a wide range of patterns encountered in neurological ICU patients. Performance was comparable or superior to that achieved by human experts. These results demonstrate that continuous, high quality, real-time computational analysis of burst suppression EEG patterns, critical to enabling optimal use of cEEG data in the neurological ICU setting, is technically achievable. Examples of clinical applications of this capacity include tracking recovery from insults such as anoxic brain injury, and EEG-based closed loop control of anesthetic delivery for the treatment of refractory status epilepticus.