Real-Time Non-Invasive Imaging and Detection of Spreading Depolarizations through EEG: An Ultra-Light Explainable Deep Learning Approach

Yinzhe Wu, Sharon Jewell, Xiaodan Xing, Yang Nan, Anthony J. Strong, Guang Yang, Martyn G. BoutelleIEEE JBHI 2024 · OpenReview

TL;DR

This paper introduces a novel, ultra-light deep learning model for real-time, non-invasive detection of Spreading Depolarizations (SDs) using EEG. By combining EEG spectrograms with temporal power vectors, the model achieves high accuracy and significantly faster processing times than previous methods, making it suitable for bedside clinical use.

What Problem Does This Paper Address?

Secondary brain injury is a major concern in neurocritical care, and Spreading Depolarizations (SDs) are a key contributor. Traditionally, SDs are detected invasively using electrocorticography (ECoG) at high sampling rates. While scalp EEG offers a non-invasive alternative, its signals are noisy and attenuated, making SD detection challenging. Existing methods often rely on fixed, high-density electrode setups and struggle with fluctuating signal power, leading to false positives. This paper aims to overcome these limitations by developing a more robust, flexible, and faster non-invasive SD detection method using EEG.

Methods

The core of this work is a novel "ultra-light-weight multi-modal deep-learning network." The approach transforms 1D EEG signals into two representations:

  1. 2D Spectrogram Imaging: EEG data is processed using Short-Time Fourier Transform (STFT) to create spectrograms. This adds a frequency dimension, allowing the model to capture subtle SD features that might be masked in traditional time-domain analysis.
  2. 1D Temporal Power Vectors: This is the more conventional approach, focusing on the signal's power over time.

These two representations are then fed into separate convolutional neural network (CNN) pathways (2D CNN for spectrograms, 1D CNN for power vectors). The features extracted from these pathways are fused using a channel-wise concatenation and a subsequent CNN layer. This dual-path approach allows the model to leverage both the rich frequency information from spectrograms and the temporal dynamics from power vectors. The model is designed to be lightweight, enabling real-time inference on CPUs without the need for high-end GPUs.

Datasets

The study utilized EEG data from two patients with traumatic brain injury (TBI) and one patient with malignant hemispheric stroke (MHS). For training and testing, the researchers augmented these limited real-world datasets. They generated simulated SD-carrying EEG segments by adding SD-specific power drop profiles to non-SD EEG tracings, incorporating realistic noise and varying AC power offsets. This augmentation strategy aimed to create a more robust dataset that reflects diverse clinical scenarios.

Key Insights and Ideas

  • Spectrograms as a Biomarker: The study highlights that EEG spectrograms, by incorporating the frequency dimension, can serve as a powerful, independent biomarker for SD detection, outperforming traditional power vector analysis in several metrics.
  • Fusion for Superiority: Combining spectrogram imaging with temporal power vectors significantly enhances SD detection accuracy and specificity compared to using either input alone.
  • Ultra-Lightweight for Real-Time: The development of an "ultra-light" deep learning model is crucial for enabling real-time, bedside monitoring on standard clinical hardware (CPUs), a significant leap from previous GPU-dependent or slow methods.
  • Single-Electrode Flexibility: The model's ability to perform accurate SD detection from a single electrode, without requiring a fixed high-density map, offers unprecedented flexibility for clinical application, especially in challenging post-surgical scenarios.
  • Explainability through Confidence Scores: Providing confidence scores alongside binary detection outcomes enhances model trustworthiness and aids clinical decision-making.

Key Results

  • High Accuracy: The proposed dual-path model achieved high specificity (0.9155), sensitivity (0.7095), and accuracy (0.8567) for SD detection on augmented EEG data.
  • Exceptional Speed: The model boasts an ultra-fast processing speed of less than 0.3 seconds per hour of EEG when run on a CPU, a dramatic improvement over conventional methods that can take hours.
  • Reduced Monitoring Time: The model requires only a 30-minute monitoring window for a decision, compared to the 30-hour requirement of some previous methods.
  • Improved Specificity: When tested on non-SD EEG data with brain injury, the model demonstrated high specificity (0.9369), indicating its ability to distinguish SDs from other signal variations.
  • Fewer False Positives: The combined model produced fewer false positive peaks in confidence scores compared to ablated models.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Limited Patient Data: The study relied on a small number of patients, and the augmented dataset, while innovative, is a simulation. Future work should involve larger, more diverse patient populations with various injury types (TBI, MHS, aSAH) and data from different electrode vendors.
  • Clinical Validation: Further validation across multiple clinical centers is needed to account for variations in practice and ensure generalizability.
  • SD Subtypes: The current model classifies all SDs together. Future research could focus on differentiating between SD subtypes (isolated, continuous, intermediate) for a more nuanced understanding.
  • Skull Integrity: While the model was tested on data where the skull was removed or altered, future work should explore its performance on patients with intact or restored skulls, potentially requiring additional fine-tuning.
  • High GCS Scores and Movement Artifacts: The current study focused on patients with low GCS scores. Future work should investigate the model's performance in patients with higher GCS scores, where voluntary movements might introduce artifacts and affect EEG tracings. More data from patients without SDs is also needed to further validate specificity.

Summary updated

Abstract

A core aim of neurocritical care is to prevent secondary brain injury. Spreading depolarizations (SDs) have been identified as an important independent cause of secondary brain injury. SDs are usually detected using invasive electrocorticography recorded at high sampling frequency. Recent pilot studies suggest a possible utility of scalp electrodes generated electroencephalogram (EEG) for non-invasive SD detection. However, noise and attenuation of EEG signals makes this detection task extremely challenging. Previous methods focus on detecting temporal power change of EEG over a fixed high-density map of scalp electrodes, which is not always clinically feasible. Having a specialized spectrogram as an input to the automatic SD detection model, this study is the first to transform SD identification problem from a detection task on a 1-D time-series wave to a task on a sequential 2-D rendered imaging. This study presented a novel ultra-light-weight multi-modal deep-learning network to fuse EEG spectrogram imaging and temporal power vectors to enhance SD identification accuracy over each single electrode, allowing flexible EEG map and paving the way for SD detection on ultra-low-density EEG with variable electrode positioning. Our proposed model has an ultra-fast processing speed (<0.3 sec). Compared to the conventional methods (2 hours), this is a huge advancement towards early SD detection and to facilitate instant brain injury prognosis. Seeing SDs with a new dimension – frequency on spectrograms, we demonstrated that such additional dimension could improve SD detection accuracy, providing preliminary evidence to support the hypothesis that SDs may show implicit features over the frequency profile.
BibTeX
@article{wu2024real,
  author = {Yinzhe Wu and Sharon Jewell and Xiaodan Xing and Yang Nan and Anthony J. Strong and Guang Yang and Martyn G. Boutelle},
  title = {{Real-Time Non-Invasive Imaging and Detection of Spreading Depolarizations through EEG: An Ultra-Light Explainable Deep Learning Approach}},
  journal = {IEEE J. Biomed. Health Informatics},
  year = {2024},
  url = {https://openreview.net/forum?id=37TBbcZ0t3},
}
Figure