Automated detection of spreading depolarizations in electrocorticography

Abstract

Abstract Spreading depolarizations (SD) in the cerebral cortex are a novel mechanism of lesion development and worse outcomes after acute brain injury, but accurate diagnosis by neurophysiology is a barrier to more widespread application in neurocritical care. Here we developed an automated method for SD detection by training machine-learning models on electrocorticography data from a 14-patient cohort that included 1,548 examples of SD direct-current waveforms as identified in expert manual scoring. As determined by leave-one-patient-out cross-validation, optimal performance was achieved with a gradient-boosting model using 30 features computed from 400-s electrocorticography segments sampled at 0.1 Hz. This model was applied to continuous electrocorticography data by generating a time series of SD probability [P SD (t)], and threshold P SD (t) values to trigger SD predictions were determined empirically. The developed algorithm was then tested on a novel dataset of 10 patients, resulting in 1,252 true positive detections (/1,953; 64% sensitivity) and 323 false positives (6.5/day). Secondary manual review of false positives showed that a majority (224, or 69%) were likely real SDs, highlighting the conservative nature of expert scoring and the utility of automation. SD detection using sparse sampling (0.1 Hz) is optimal for streaming and use in cloud computing applications for neurocritical care.