EEG Bad-Channel Detection Using Multi-Feature Thresholding and Co-Occurrence of High-Amplitude Transients

This paper introduces a publicly available MATLAB module for EEG bad-channel detection that combines multi-feature thresholding with co-occurrence-based clustering to identify and group artifactual channels through an interpretable, human-in-the-loop review process, serving as a quality-control step prior to ICA.

The Gist

Imagine you are trying to listen to a choir of 128 singers (the electrodes on an EEG cap) to understand the song they are singing (the brain's activity). Most of the time, they sing in harmony. But occasionally, one singer might have a microphone that's broken, another might be coughing, and a third might be standing too close to a noisy air conditioner.

If you don't fix these problems before you start analyzing the song, the broken microphones and coughs will ruin your understanding of the whole choir.

This paper introduces a new digital "Sound Check" tool for brainwave recordings. It's a piece of software designed to help researchers find the "broken microphones" (bad channels) in their data, but with a very human twist: it doesn't just automatically delete them; it helps the researcher make the final call.

Here is how it works, broken down into simple concepts:

1. The Problem: Why "Automatic" Isn't Always Perfect

Older tools try to be fully automatic. They look at a channel and say, "That looks weird, delete it!"

• The Flaw: Sometimes, a channel looks weird because it's broken. But sometimes, it looks weird because everyone is reacting to the same thing (like a sudden loud noise or a blink). If the computer deletes those channels automatically, it might throw away important information about how the brain reacts to that noise.
• The Goal: We need a tool that spots the troublemakers but lets a human decide if they are actually broken or just part of the group chaos.

2. The Solution: A Three-Step Detective Process

The authors created a MATLAB module (a specialized software tool) that acts like a detective with three different magnifying glasses.

Step A: The "Neighbor Check" (Is this channel out of step?)

Imagine the singers are sitting in a circle. If one singer is singing a completely different tune than the people sitting next to them, that's a red flag.

• How the tool does it: It compares every electrode to its physical neighbors. If a channel's signal looks totally different from the ones touching it, the tool marks it as "Suspicious."

Step B: The "Volume Check" (Is this channel screaming?)

Some microphones are so broken they just scream static or go completely silent.

• How the tool does it: It looks for signals that are impossibly loud (like a 1000-volt spike) or impossibly flat.
  • Super Loud? Marked as "Bad" immediately.
  • Just a bit too loud? Marked as "Suspicious."

Step C: The "Party Crashers" Check (The Secret Sauce)

This is the most creative part of the paper.
Imagine a party where everyone suddenly jumps at the same time because a car backfired outside.

• The Old Way: A computer might look at one person jumping and say, "You're crazy, leave the party!"
• The New Way: This tool notices that five people jumped at the exact same millisecond. It realizes, "Ah, they aren't crazy; they all heard the car!"
• How it works: The software looks for "high-amplitude transients" (sudden, loud spikes). It groups channels together if they spike at the exact same time.
  • If a group of channels spikes together, the tool puts them in a "Cluster."
  • It tells the researcher: "Hey, look at this group. They are all acting weird at the same time. Is it a broken wire, or is it a shared event (like a blink or a muscle twitch)?"

3. The Human-in-the-Loop: The "Review Room"

Instead of the computer deleting the data, it opens a Review Room (a graphical interface).

• It shows the researcher the "Suspicious" channels and the "Clusters" of channels that spiked together.
• The researcher can look at the waveforms and say:
  • "Okay, this cluster is just eye-blinks. I'll keep them."
  • "This one channel is totally disconnected from the group. I'll delete it."
• Why this matters: It combines the speed of a computer with the wisdom of a human. The computer does the heavy lifting of finding the trouble spots, but the human makes the final decision.

4. The Result: A Cleaner Song

Once the researcher approves the list of "Bad Channels," the software removes them.

• Before: The recording looks chaotic, with huge spikes making it hard to see the real brain waves.
• After: The "broken microphones" are gone. The remaining data is clean, and the researcher can now use advanced tools (like ICA) to separate the brain's true song from the background noise.

Summary Analogy

Think of this tool as a smart traffic cop at a busy intersection.

• Old Automated Systems: Are like a robot that instantly bans any car that speeds up, even if it's just a fire truck rushing to an emergency.
• This New Tool: Is a smart cop who sees a car speeding. Instead of banning it immediately, the cop checks: "Is this car alone? Or is it part of a parade?"
  • If it's alone and speeding? Banned.
  • If it's part of a parade (a cluster of similar events)? Let it pass, but flag it for review.

The paper concludes that by using this "Multi-Feature Thresholding" and "Co-Occurrence Clustering," researchers can clean their brain data more accurately, ensuring they don't accidentally throw away valuable information while fixing the broken parts.

Technical Summary

1. Problem Statement

Electroencephalography (EEG) recordings are susceptible to non-physiological contamination from electrode-scalp coupling issues, motion, and hardware interference. In high-density montages, these failures can be persistent (poor contact) or intermittent (transient bursts).

• The Challenge: Existing automated pipelines (e.g., FASTER, PREP, EEGLAB) often rely on global statistics or single-channel thresholds. They struggle to distinguish between:
  1. True channel failure: A specific sensor malfunctioning.
  2. Shared artifacts: Physiological or environmental events (e.g., eye blinks, muscle bursts, cable movement) that affect multiple channels simultaneously.
• The Risk: Fully automated rejection risks discarding valid physiological data (by removing channels involved in shared artifacts) or failing to catch intermittent failures that do not meet strict global thresholds. Furthermore, automated decisions often lack interpretability, making it difficult for researchers to justify channel removal in study-specific contexts.

2. Methodology

The authors introduce a MATLAB Module (SENSI EEG PREPROC Bad-Channel Detection) designed as a quality-control step before Independent Component Analysis (ICA). The workflow emphasizes human-in-the-loop validation rather than fully automated rejection.

The methodology consists of three main stages:

A. Multi-Feature Suspiciousness Scoring

The module computes complementary features to pre-label channels as Good, Suspicious, or Bad.

1. Robust Standardization: Signals are standardized using the Median and Median Absolute Deviation (MAD) to stabilize scaling against large transients.
2. Neighbor Dissimilarity: For each time window, Pearson correlations are computed between a channel and its spatial neighbors (defined by montage-specific maps). The median correlation is converted to a dissimilarity score. High dissimilarity flags a channel as suspicious.
3. Amplitude Screening:
   • Hard Cutoff: Channels with maximum absolute amplitude exceeding a threshold (e.g., 1000 µV) are pre-labeled Bad.
   • Soft Warning: Channels with unusually high log-transformed amplitudes (Z-score $\ge$ 2) are pre-labeled Suspicious.
4. Variance Analysis:
   • Global Variance: Detects unusually noisy (high variance) or flat/poorly conductive (low variance) channels.
   • Temporal Variability: Detects channels where variance fluctuates significantly over time (instability).

B. Clustering Based on Co-Occurrence of Transients

To identify channels affected by shared artifact sources, the module groups channels based on the timing of high-amplitude events.

1. Transient Detection: Robustly standardized signals are thresholded (e.g., $|Z| \ge 14$) to create binary masks of extreme events.
2. Max-Pooling: These masks are downsampled into short time blocks (e.g., 200ms) to create sparse event-timing vectors.
3. Jaccard-like Distance: A distance metric is calculated between channel pairs based on the overlap of their transient events.
   • $D = 1 - \frac{\text{Intersection}}{\max(\text{Union})}$
   • Crucially, this metric focuses only on the intersection of extreme events, ignoring periods where both channels are quiet. This isolates shared artifactual structure from shared neural activity.
4. Clustering: Channels are grouped into connected components if their distance is below a threshold ($\epsilon$).
   • Interpretation: Small clusters (few channels) likely indicate channel-specific failures. Large clusters often indicate global artifacts (e.g., eye blinks).
   • EOG Handling: Clusters overlapping with EOG channels are constrained to prevent automatic rejection of ocular artifacts.

C. Interactive Review Interface

The module does not enforce automatic rejection. Instead, it launches a GUI (reviewBadChsUI) where users:

• View channels organized by their transient-event clusters.
• Inspect time-domain overlays of pre-flagged channels.
• Toggle channel status between Good (0), Suspicious (1), and Bad (2).
• Final Output: Only channels explicitly marked as Bad (2) are returned in the final list. "Suspicious" channels are retained for further review but not removed.

3. Key Contributions

• Relational Structure over Isolation: Unlike traditional methods that evaluate channels in isolation, this approach uses event-timing similarity to group channels. This allows researchers to see if a "bad" channel is part of a larger artifact pattern (suggesting it should be kept for ICA) or an isolated failure (suggesting removal).
• Human-in-the-Loop Design: The system shifts the burden from "automated decision-making" to "automated prioritization." It reduces manual inspection time by grouping related channels and highlighting outliers, but leaves the final classification to the expert.
• Interpretability and Transparency: The module provides visual diagnostics (cluster graphs, distance matrices, feature score plots) and an audit trail of decisions, addressing the "black box" nature of many automated pipelines.
• Open Source Implementation: The code is released as a publicly available, documented MATLAB module with example datasets and workflows.

4. Results (Illustrative Analysis)

The paper demonstrates the workflow on a 128-channel EEG dataset:

• Clustering Visualization: The force-directed graph successfully separated channels into clusters. Large clusters (e.g., C1) represented channels with no transient events (likely good), while small, isolated clusters were flagged for review.
• Feature Scoring: The feature-score summary plot effectively triaged channels, highlighting specific outliers in amplitude and neighbor dissimilarity that required attention.
• Interactive Review: The GUI allowed the user to rapidly validate pre-flagged channels. In the example, 7 channels were marked as "Bad" out of 128.
• Before/After Comparison: Removing the 7 bad channels reduced the amplitude scale of the overlay plot from $\pm 20,000$ µV to $\pm 1,000$ µV, demonstrating the removal of extreme excursions that would otherwise dominate the data visualization and distort downstream analysis.

5. Significance

• Preservation of Data: By distinguishing between shared artifacts and sensor failure, the method prevents the premature removal of channels that contain valid physiological signals, thereby preserving the covariance structure necessary for effective ICA decomposition.
• Reproducibility: The interactive nature ensures that channel removal decisions are explicit, documented, and justifiable, which is critical for high-impact neuroscientific research.
• Workflow Integration: It fills a gap in the preprocessing pipeline between basic filtering and ICA, offering a robust, interpretable step that complements existing automated tools rather than replacing them.
• Adaptability: While currently supporting specific montages (EGI, BioSemi, Easycap), the modular design allows users to define their own neighbor maps, making it adaptable to various EEG setups.

In summary, this paper presents a pragmatic, transparent, and scientifically rigorous approach to EEG bad-channel detection that balances automation with expert oversight, specifically addressing the complexities of high-density EEG data where artifacts are often spatially structured.