sEEGnal: an automated EEG preprocessing pipeline evaluated against expert-driven preprocessing

The paper introduces sEEGnal, a fully automated and modular EEG preprocessing pipeline that achieves performance comparable to expert-driven methods while offering greater consistency, scalability, and reproducibility through its BIDS-based framework.

The Gist

Imagine you are trying to listen to a beautiful symphony orchestra, but the recording is filled with static, coughing, people walking by, and even the sound of a car engine passing outside. To hear the music clearly, you need a sound engineer to clean up the recording, removing the noise while keeping the music intact.

In the world of brain science, EEG (Electroencephalography) is that recording. It captures the brain's electrical "music," but it's often messy with "noise" like eye blinks, muscle tension, or bad wires.

This paper introduces sEEGnal, a new, fully automated "sound engineer" for brain data. Here's the story of how it works and why it matters, explained simply.

The Problem: The "Human Editor" Bottleneck

For years, cleaning up brain data has been like hiring a team of highly skilled, expensive sound editors.

• The Good: These human experts are great. They know exactly what to cut and what to keep.
• The Bad: They are slow, they get tired, and they all have slightly different opinions. If you have 1,000 recordings, it would take a human team months to clean them up. Plus, if you ask two different experts to clean the same recording, they might do it slightly differently, making it hard to compare results later.

The Solution: Enter sEEGnal

The authors built sEEGnal, a robot sound engineer. It's a computer program that does the same job as the human experts but does it in seconds, never gets tired, and always follows the exact same rules.

Think of sEEGnal as a smart, modular kitchen robot that prepares a meal. It doesn't just throw everything in a pot; it has specific stations:

1. The Standardizer: It organizes the ingredients (data) into neat, labeled jars (using a system called BIDS) so anyone can find them later.
2. The Bad Channel Detective: It checks every microphone (electrode) on the head. If a microphone is broken, covered in jelly (gel bridges), or just making weird static, it marks it as "bad" and ignores it.
3. The Noise Filter: It listens for specific types of noise:
   • Eye blinks: "Oh, someone blinked! That's not brain music."
   • Muscle tension: "That's a jaw clench, not a thought."
   • Sensor pops: "That's a loose wire, not a brain signal."
   • It uses a clever trick (called ICA) to separate the "brain music" from the "noise instruments" and throws the noise away.

The Big Test: Robot vs. Human

The big question was: Can the robot do as good a job as the human experts?

To find out, the researchers took brain recordings from two different groups of people (one public group, one private group) and did two things:

1. They had human experts clean the data the old-fashioned way.
2. They let sEEGnal clean the exact same data.

Then, they compared the results like a judge at a cooking competition.

The Results:

• The Taste Test (Brain Signals): When they looked at the final "music" (the brain waves), the robot's version sounded almost identical to the human's version. The patterns of brain activity were preserved perfectly.
• The Consistency: This is where the robot won big. If you ask 10 humans to clean the same data, they might all do it slightly differently. But if you ask the robot to do it 10 times, it does it exactly the same way every single time. It's perfectly consistent.
• The Speed: The robot was just as fast as the humans (sometimes faster) and used less computer memory.

Why This Matters

Imagine you are a scientist trying to find a cure for a disease. You need to analyze brain data from thousands of patients.

• Before sEEGnal: You'd need a huge team of experts working for years, and their results might be inconsistent.
• With sEEGnal: You can process thousands of patients in a day, with the same high quality as a top expert, and you know the results are reliable every time.

The Bottom Line

sEEGnal is a game-changer. It's an open-source tool (free for everyone to use) that automates the boring, difficult, and time-consuming part of brain research. It proves that a computer can be just as good as a human expert at cleaning brain data, but it's faster, cheaper, and never has a bad day.

It's like upgrading from a hand-cranked record player to a high-tech digital streaming service: the music is still the same, but the experience is smoother, faster, and available to everyone.

Technical Summary

1. Problem Statement

Electroencephalography (EEG) preprocessing is a critical but labor-intensive step required to remove physiological (e.g., eye blinks, muscle activity) and non-physiological (e.g., electrode noise, power line interference) artifacts. Current best practices rely on semi-automatic pipelines supervised by human experts. This approach presents significant limitations:

• Scalability: Manual supervision is time-consuming and unaffordable for large-scale datasets.
• Reproducibility: Outcomes vary significantly between different experts (inter-rater variability).
• Accessibility: Most existing tools require programming skills (MATLAB/Python), limiting their use to specialized researchers.
• Lack of Unified Framework: Existing automated tools often prioritize either standardized workflows or data-driven detection but fail to integrate both into a single, interpretable framework.

There is a need for a fully automated, modular, and open-source pipeline that matches the performance of human experts while ensuring consistency and computational efficiency.

2. Methodology: The sEEGnal Pipeline

sEEGnal is a fully automated, modular pipeline written primarily in Python (with C++ modules for performance) built upon the MNE-Python framework. It is designed to be configurable via JSON files and adheres to the Brain Imaging Data Structure (BIDS) standard for data interoperability.

The pipeline consists of three core modules:

A. Standardization Module

• Converts raw EEG data into the BIDS-compliant format, ensuring structured folder structures, filenames, and metadata for reproducibility and interoperability across different analysis frameworks.

B. Bad Channel Detection Module

This module identifies unreliable electrodes based on specific physiological and technical criteria. It uses a stepwise approach where channels marked as "bad" are excluded from subsequent checks. Criteria include:

1. High Impedance: Channels exceeding a user-defined threshold (default: 200 kΩ).
2. Impossible Amplitude: Channels with standard deviations outside physiological ranges (default: <1 µV or >500 µV).
3. Power Spectrum Anomalies: Detection of abnormal power in the 45–55 Hz range (indicating line noise or hardware issues).
4. Gel Bridges: Identification of adjacent channels (<5 cm) with correlation coefficients >0.999, indicating conductive gel bridging.
5. High Amplitude Variance: Channels with variance 10x greater than the recording average.

• Pre-processing for detection: Data is bandpass filtered (1–100 Hz), downsampled (500 Hz), notched (50 Hz), and decomposed using Second Order Blind Identification (SOBI) and classified via ICLabel to assist in context.

C. Artifact Detection Module

This module identifies the onset and duration of artifacts using a two-pass Independent Component Analysis (ICA) strategy to prevent dominant artifacts from skewing component estimation.

• Pass 1: Decomposes data and detects Muscle (EMG), Sensor (jumps/pops), and Other artifacts.
  • Muscle: Reconstructs time series from ICs labeled "muscle" by ICLabel (certainty >0.7) and detects high-frequency peaks (110–145 Hz).
  • Sensor: Detects abrupt steps in the 0.5–5 Hz band (5x standard deviation).
  • Other: Detects impossible high amplitudes (>500 µV) in the 2–45 Hz band.
• Pass 2: Re-decomposes data after removing epochs with the artifacts found in Pass 1.
• Pass 3: Detects Eye-related (EOG) artifacts by analyzing frontal vs. background channel standard deviations in the 1–5 Hz band.

3. Evaluation Strategy

Since no "ground truth" exists for real-world EEG, the authors evaluated sEEGnal against manual preprocessing performed by multiple EEG experts.

• Datasets:
  • LEMON: Public dataset (20 subjects, resting state, eyes open/closed).
  • AI-Mind: Private dataset (20 subjects, MCI continuum, resting state).
  • Test-Retest: Subset of AI-Mind (20 subjects, repeated recordings in the same session) to assess stability.
• Metrics:
  1. Preprocessing Metadata: Time/memory consumption, number of bad channels, artifact duration, and rejected Independent Components (ICs).
  2. Neurophysiological Measures:
     • Power Spectral Density (PSD): Calculated via Welch's method.
     • Functional Connectivity: Estimated using Phase Locking Value (PLV) across standard frequency bands (Delta to Gamma).
  3. Statistical Analysis: Paired t-tests for metadata and Normalized Root Mean Squared Error (NRMSE) to quantify similarity between sEEGnal and expert outputs.

4. Key Results

A. Metadata and Efficiency

• Performance: sEEGnal detected a comparable number of bad channels, artifacts, and rejected components as human experts (no significant statistical difference).
• Consistency: sEEGnal showed significantly lower variability in processing time compared to experts (experts had high standard deviations due to human factors).
• Resource Usage: sEEGnal consumed significantly less memory than the expert workflow.
• Artifact Classification: While total artifact counts were similar, the types differed. sEEGnal identified proportionally more eye and muscle artifacts, while experts identified more "electronic jumps" and "other" artifacts.

B. Neurophysiological Measures (PSD and PLV)

• Power Spectra: sEEGnal results showed high resemblance to expert results.
  • NRMSE: All frequency bands had NRMSE < 0.08 (differences < 8% of metric range).
  • Spatial Bias: Largest differences occurred in frontal channels within the Delta band, likely due to eye-movement artifact handling.
• Functional Connectivity (PLV):
  • NRMSE: Values were generally < 0.17 (Delta band showed highest deviation).
  • Observation: sEEGnal tended to produce slightly lower PLV values than experts, particularly in frontal channels, suggesting a more conservative approach to artifact removal that might slightly dampen connectivity estimates.
• Test-Retest Stability: sEEGnal demonstrated exceptional stability.
  • PSD NRMSE: < 0.045 (4.5% difference) between repeated recordings.
  • PLV NRMSE: < 0.09 (9% difference).
  • This indicates the pipeline is highly reproducible, unlike human experts whose decisions can vary between sessions.

5. Key Contributions

1. Fully Automated Pipeline: A modular, end-to-end solution that requires no programming knowledge from the end-user.
2. Expert-Level Performance: Validated against human experts, demonstrating comparable accuracy in identifying bad channels and artifacts while preserving key neurophysiological features.
3. Enhanced Reproducibility: Eliminates inter-rater variability, providing a consistent standard for large-scale studies.
4. BIDS Compliance: Built-in support for the BIDS standard, facilitating data sharing and multi-center collaboration.
5. Open Source: The code is publicly available (GitHub), promoting transparency and community adoption.

6. Significance and Limitations

Significance:
sEEGnal addresses the bottleneck of manual EEG preprocessing, enabling the processing of large-scale datasets (e.g., clinical trials, big data neuroscience) with a level of consistency and efficiency unattainable by human teams. It serves as a robust "gold standard" alternative for studies where manual review is impractical.

Limitations:

• Dataset Scope: Validation was limited to two datasets (one public, one private) and resting-state data. Task-based EEG performance was not evaluated.
• Dependency: The pipeline relies heavily on external packages, specifically ICLabel, which is trained on millions of expert-labeled recordings. While robust, this creates a dependency on the external tool's updates.
• Expert Count: The number of human experts used for comparison was relatively small.

Conclusion:
sEEGnal represents a state-of-the-art solution for automated EEG preprocessing. It successfully balances automation with the preservation of neurophysiological integrity, offering a scalable, reproducible, and efficient tool for the neuroscience community.