Neural Noise

Automated detection of spreading depolarizations in electrocorticography

2 min read

  • Automated algorithms can detect spreading depolarizations from ECoG signals.1
  • ECoG-based spreading-depolarization detection supports ICU monitoring in acute brain injury.1
  • The approach is algorithm-ready for integration into closed-loop BCI and electrophysiology pipelines.1 1

Weekly enrichment (2026-07-20)

  • The Scientific Reports (2025) study trained machine-learning models on ECoG from a 14-patient development cohort (Cohort 1), which contained 1,548 expert-scored SD direct-current waveforms across 1,128 hours of recording, with SDs present in 11 of 14 patients.23
  • Optimal performance came from a gradient-boosting model using 30 features computed from 400-second ECoG segments down-sampled to 0.1 Hz, selected via leave-one-patient-out cross-validation.23
  • The model outputs a continuous SD-probability time series P_SD(t); SD predictions are triggered when both a probability threshold and a minimum-duration threshold are empirically exceeded.3
  • On a naive 10-patient validation cohort (Cohort 2; 1,187 hours, 1,949-1,953 scored SDs in 8 of 10 patients), the algorithm achieved 1,252 true-positive detections (~64% sensitivity) with 323 false positives (~6.5 per day).23
  • Secondary manual review found 224 of the 323 false positives (69%) were likely real SDs, indicating expert scoring is conservative and that automation may recover missed events.23
  • Sparse 0.1 Hz sampling makes the method well suited to streaming and cloud-computing pipelines for neurocritical-care monitoring.23
  • Clinically, SDs occur in roughly 60% of patients after severe traumatic brain injury (and at higher incidence after subarachnoid hemorrhage and malignant hemispheric stroke) and are independently associated with worse outcomes, motivating automated bedside detection.4
  • An earlier prototype (Neurocritical Care, 2021) validated real-time bedside SD detection on 91 ~24-hour files from 18 patients, showing a software-vs-manual count regression slope of 0.786 (R^2 = 0.842) with ~79% relative sensitivity but preserved detection of dense SD clusters.4
  • Newer approaches push further: a multi-scale signal-image Transformer-CNN fusion model on 500 h of neuro-ICU ECoG reported 92.6% accuracy and 84.9% sensitivity, detecting SD onset on average 8 minutes before expert observation, while a WAVEFRONT-based method demonstrated noninvasive SD detection from 19-electrode scalp EEG against intracranial ground truth.56

Footnotes

  1. https://news.google.com/rss/articles/CBMiX0FVX3lxTE1uMXptSGVyNGFWUGRsR3pFM0JnUGtELWdKa3NPWE9yeDRVaENDRmVTc0h1b1ZoQTdIY3Rhb212VG05QzBVUndGemFVYUtaQWZSS1JaaS1La3c2Q0dLak1B?oc=5 ↩ ↩2 ↩3 ↩4

  2. https://www.nature.com/articles/s41598-025-91623-7 ↩ ↩2 ↩3 ↩4 ↩5

  3. https://preview-www.nature.com/articles/s41598-025-91623-7 ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  4. https://pmc.ncbi.nlm.nih.gov/articles/PMC8536628/ ↩ ↩2

  5. https://doi.org/10.1109/tbme.2026.3673475 ↩

  6. https://pmc.ncbi.nlm.nih.gov/articles/PMC10439895/ ↩

Graph View