Automated detection of spreading depolarizations in electrocorticography

Area of Science:

• Neuroscience
• Computational Neuroscience
• Medical Technology

Background:

• Spreading depolarizations (SD) are implicated in lesion development and poor outcomes following acute brain injury.
• Accurate neurophysiological diagnosis of SD is currently a significant challenge in neurocritical care.

Purpose of the Study:

• To develop and validate an automated method for detecting spreading depolarizations (SD) in electrocorticography (ECoG) data.
• To enable more widespread clinical application of SD detection in neurocritical care settings.

Main Methods:

• Machine learning models were trained on ECoG data from 14 patients with 1,548 expert-identified SD events.
• A gradient-boosting model utilizing 30 features from 400-second ECoG segments sampled at 0.1 Hz demonstrated optimal performance.
• The algorithm generated a time series of SD probability [P_SD(t)] applied to continuous ECoG data.

Main Results:

• The automated method achieved 64% sensitivity (1,252 true positives out of 1,953) with 323 false positives per day on a novel patient dataset.
• A secondary review indicated that 69% of false positives were likely actual SDs, suggesting a conservative expert scoring approach.
• Sparse sampling (0.1 Hz) proved optimal for real-time streaming and cloud-based applications.

Conclusions:

• An automated machine learning algorithm effectively detects spreading depolarizations (SD) in neurocritical care.
• The developed method, utilizing sparse sampling, is suitable for real-time streaming and cloud computing.
• Automation of SD detection can overcome current diagnostic barriers, facilitating broader clinical use.