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

Technical Summary: Real-Time Non-Invasive Imaging and Detection of Spreading Depolarizations through EEG

Problem Statement
Spreading depolarizations (SDs) are a critical cause of secondary brain injury in neurocritical care, yet their detection currently relies on invasive electrocorticography (ECoG). While non-invasive electroencephalogram (EEG) offers a potential alternative, detecting SDs via scalp EEG is hindered by signal attenuation, noise, and the limitations of conventional analysis methods. Existing approaches typically analyze 1-D temporal power changes over fixed, high-density electrode maps. These methods often fail in clinical settings due to fluctuating EEG AC power offsets, the impracticality of fixed high-density electrode setups in post-surgical trauma patients, and the inability to distinguish SD-specific power suppression from other non-pathological signal variations. Furthermore, previous algorithms lack real-time processing capabilities on standard bedside hardware (CPUs) and often require extensive data windows (e.g., 30 hours) to make decisions, delaying prognosis.

Methodology
The authors propose a novel, ultra-lightweight, multi-modal deep learning framework that transforms SD detection from a 1-D time-series task into a sequential 2-D imaging task. The methodology proceeds through the following stages:

1. Signal Preprocessing and Feature Extraction:

   • EEG signals are bandpass-filtered (0.5–45 Hz) to remove DC drift and high-frequency noise.
   • Spectrogram Generation: Using Short-Time Fourier Transform (STFT) with a Hanning window, the AC EEG signal is converted into 2-D spectrogram images (30 × 30 arrays) with a time resolution of 1 minute and frequency resolution of 1 Hz. The analysis focuses on the delta band (0.5–45 Hz), specifically extracting power arrays for 0.5–1.85 Hz where signal power is highest.
   • Power Vector Generation: Simultaneously, 1-D temporal power vectors are calculated for the same 30-minute sliding windows.

2. Network Architecture:

   • The model employs a dual-path Convolutional Neural Network (CNN) architecture.
   • Image Path: Processes the 2-D spectrogram images through three blocks of 2-D convolutional layers (3×3 kernels), ReLU activations, and Max Pooling.
   • Vector Path: Processes the 1-D power vectors through three blocks of 1-D convolutional layers.
   • Fusion: The feature maps from both paths are reshaped, concatenated, and passed through a final CNN block to fuse spectrogram imaging features with temporal power vectors.
   • Output: The network produces a binary outcome (SD present/absent) for each 30-minute window.

3. Confidence Scoring:

   • To enhance explainability, a 30-minute sliding window sums the binary outcomes around each time point, generating a confidence score (0–30). This highlights the peak of SD depression and provides a quantitative measure of detection certainty.

4. Data Strategy:

   • Due to the scarcity of co-registered EEG and ECoG data, the study utilized data from two patients (one TBI with 17 confirmed SDs, one MHS with no SDs).
   • To train the model, the authors simulated a large dataset (1,700 hours for training and testing) by augmenting non-SD EEG tracings with SD-specific power suppression profiles and varying levels of noise, validated against COSBID consensus guidance.

Key Contributions

1. Spectrogram as a Biomarker: The study introduces the use of frequency-restricted 2-D spectrogram imaging as a standalone and superior biomarker for SD detection compared to traditional 1-D power vectors, demonstrating that the frequency domain contains implicit features of SDs.
2. Ultra-Lightweight Real-Time Processing: The model is designed for CPU-only execution, achieving an inference time of <0.3 seconds per hour of EEG (including preprocessing). This contrasts sharply with previous methods requiring hours of processing or GPU hardware.
3. Flexible Electrode Configuration: The framework operates on a single electrode basis, removing the requirement for fixed, high-density electrode maps. This allows for flexible electrode positioning suitable for diverse clinical scenarios, including post-craniotomy patients.
4. Reduced Monitoring Window: The model requires only a 30-minute window for decision-making, significantly reducing monitoring pressure compared to the 30-hour windows required by conventional methods.
5. Explainability: The generation of confidence scores provides clinicians with a quantitative rationale for detection decisions, addressing the "black box" nature of previous deep learning approaches.

Results

• Performance on Augmented SD Data: The dual-path model achieved a specificity of 0.9155, sensitivity of 0.7095, and accuracy of 0.8567 in binary detection (Stage 5). In terms of peak detection (Stage 6), it achieved a sensitivity of 0.9848 with the fewest false positive peaks compared to ablated models (spectrogram-only or power-vector-only).
• Specificity on Non-SD Data: When tested on EEG from a patient with brain injury but no SDs, the dual-path model demonstrated the highest specificity (0.9369) and the lowest Euclidean distance to expected confidence scores, effectively differentiating SD-specific changes from general signal fluctuations.
• Ablation Studies: The spectrogram-only path showed higher sensitivity (0.8325) than the power-vector-only path (0.6175) but lower specificity. The fusion of both modalities provided the best balance, confirming the utility of the spectrogram as an independent feature source while the power vector aids in specificity.
• Inference Speed: On a standard CPU (Intel Xeon 2.20 GHz), the total processing time was approximately 0.21 seconds per hour of EEG (0.13s inference + 0.08s preprocessing). On a low-specification GPU (NVIDIA T4), inference took only 4 milliseconds per hour.

Significance and Claims
The paper claims to present the first real-time, non-invasive SD detection framework capable of operating on standard bedside CPUs without GPU acceleration. By shifting the detection paradigm from 1-D time-series analysis to 2-D spectrogram imaging, the study provides preliminary evidence that frequency-domain features improve detection accuracy. The authors assert that their method facilitates early SD detection and instant brain injury prognosis by drastically reducing processing time (from hours to seconds) and the required data window (from 30 hours to 30 minutes). Furthermore, the single-electrode approach offers a pathway for SD monitoring in ultra-low-density EEG setups with variable electrode positioning, potentially expanding access to SD screening for millions of stroke and trauma patients who cannot undergo invasive ECoG monitoring. The authors note that while the current data includes patients with craniectomies (which may enhance signal transmission), the method is expected to be transferable to patients with intact skulls, pending further validation on larger, more diverse datasets.