Automated EEG-based prediction of delayed cerebral ischemia after subarachnoid hemorrhage

Abstract

This dataset supports our study on developing an automated algorithm for predicting delayed cerebral ischemia (DCI) after subarachnoid hemorrhage (SAH) using electroencephalography (EEG). We provide EEG recordings and clinical data from 113 moderate to severe grade SAH patients, along with the code for our machine learning model that integrates multiple EEG features for DCI prediction. The algorithm achieves an area under the receiver-operator curve of 0.73 by day 5 after SAH with good calibration between 48-72 hours, demonstrating the potential of automated, multi-featured EEG assessment for DCI risk prediction.

The citation is: 

Zheng WL, Kim JA, Elmer J, Zafar SF, Ghanta M, Moura Junior V, Patel A, Rosenthal E, Brandon Westover M. Automated EEG-based prediction of delayed cerebral ischemia after subarachnoid hemorrhage. Clin Neurophysiol. 2022 Nov;143:97-106. doi: 10.1016/j.clinph.2022.08.023. Epub 2022 Sep 11. PMID: 36182752; PMCID: PMC9847346.

Code is here: https://github.com/bdsp-core/SAH_DCI_Prediction_EEG

Background

Delayed cerebral ischemia (DCI) is a major cause of morbidity following aneurysmal subarachnoid hemorrhage (SAH). While vasospasm is the best-known risk factor for DCI, multiple mechanisms contribute, including cortical spreading depolarization, impaired cerebral autoregulation, microcirculatory dysfunction, neuroinflammation, and microthrombosis. Continuous EEG (cEEG) has emerged as a promising diagnostic tool for broader DCI detection, with several EEG features showing association with increased DCI risk. However, previous studies required manual EEG review by neurophysiologists and/or manual identification of artifact-free EEG segments, limiting the clinical implementation of cEEG as a DCI prediction tool. Our automated approach addresses this limitation by combining a weighted artifact detection calculation with multiple auto-calculated quantitative cEEG features.

Methods

We included patients with moderate to severe grade SAH (Hunt Hess 4-5 or Fisher grade 3) treated at Massachusetts General Hospital between September 2011 and January 2015. Continuous EEG monitoring was performed for ischemia detection as part of routine clinical care, typically between days 2-10 after SAH. We excluded patients who developed status epilepticus and those with DCI events occurring >24 hours after discontinuation of cEEG monitoring.

EEG data was recorded with 19 monopolar channels according to the international 10-20 system, filtered between 0.5-30 Hz with a 60 Hz notch filter, resampled to 200 Hz, and rereferenced to create 18 bipolar channels. We developed a weighting approach based on signal quality to reduce contributions of noise and artifacts. For each 5-minute epoch, we extracted 9 EEG features: epileptiform discharge (ED) burden, Shannon entropy, delta, theta, alpha, beta, and total band power, alpha/delta ratio (ADR), and percent alpha variability (PAV).

We developed a time-sensitive "max carry forward" machine learning model using Random Forest classifiers to predict DCI every six hours after SAH onset. Features were derived from whole-brain averages, hemispheric measurements, and vascular territory calculations. Model performance was evaluated using area under the receiver operating characteristic curve (AUC-ROC) and calibration error.

Data Description

The dataset contains:

1. Deidentified EEG recordings from 113 SAH patients (58 with DCI, 49 without)
2. Clinical and demographic information including:
   • Age and sex
   • SAH grade (Hunt-Hess and Fisher scores)
   • Timing of SAH onset and DCI events
   • Patient outcomes
3. Extracted EEG features including:
   • Epileptiform discharge burden
   • Shannon entropy
   • Spectral power bands (delta, theta, alpha, beta, total)
   • Alpha/delta ratio (ADR)
   • Percent alpha variability (PAV)
   • Vascular territory-specific measurements
   • Hemispheric asymmetry calculations
4. Code for artifact detection, feature extraction, and the machine learning prediction model

Usage Notes

Data and code to generate all results and figures from the publication are provided here. The repository contains the complete codebase needed to reproduce our analysis, including EEG preprocessing, feature extraction, machine learning model implementation, and visualization code. All figures in the publication can be reproduced using the provided scripts.

Ethics

This study was approved by the Massachusetts General Hospital institutional review board (IRB #2013P001024).

Acknowledgements

JAK received funding from NINDS (R25N065743, K23NS112596-01A1), American Heart Association, and Bee Foundation. JE received funding from NINDS (5K23NS097629). MBW received funding from Glenn Foundation for Medical Research, American Federation for Aging Research (Breakthroughs in Gerontology), the American Academy of Sleep Medicine Strategic Research Award, DoD Moberg ICU Solutions, Inc. subcontract, and NIH (1R01NS102190, 1R01NS102574, 1R01NS107291, 1RF1AG064312). WLZ was supported by NIH (1R01NS102574). SFZ is supported by NIH (K23NS114201). 

Conflicts of Interest

M.B.W. is a co-founder of and holds equity in Beacon Biosignals. Beacon Biosignals did not contribute funding nor played any role in the study. The authors declare that they have no other known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References

1. Zheng WL, Kim JA, Elmer J, Zafar SF, Ghanta M, Moura Junior V, Patel A, Rosenthal E, Brandon Westover M. Automated EEG-based prediction of delayed cerebral ischemia after subarachnoid hemorrhage. Clin Neurophysiol. 2022 Nov;143:97-106. doi: 10.1016/j.clinph.2022.08.023. Epub 2022 Sep 11. PMID: 36182752; PMCID: PMC9847346.