Semi-Automated Identification of EKG and Trigger Artifacts in EEG Using ICA and Spectral Characteristics

This paper introduces "SENSI-EEG-Preproc-ICA-EKG-Trigger," a semi-automated MATLAB module that accelerates and standardizes the identification of EKG and trigger artifacts in EEG data by leveraging spectral characteristics to prioritize candidate components for efficient, user-verified manual review.

The Gist

Imagine you are trying to listen to a quiet conversation in a crowded room. You have a high-tech microphone (the EEG machine) that records every sound, but the recording is messy. It's filled with the hum of the air conditioner, the clinking of silverware, and the sudden "beep" of a timer.

Your goal is to isolate the conversation (the brain waves) and remove the noise. To do this, scientists use a clever trick called ICA (Independent Component Analysis). Think of ICA as a magical audio mixer that takes the messy recording and splits it into 64 separate tracks (one for each electrode on the head). Now, instead of one noisy song, you have 64 individual instrument tracks.

The Problem:
The problem is that you don't know which track is the "conversation" and which track is the "silverware clinking." Usually, a human has to sit down and listen to all 64 tracks one by one, decide which ones are garbage, and delete them. This is slow, boring, and everyone does it differently.

The Solution (The Paper):
This paper introduces a new tool called SENSI-EEG-Preproc. Think of it as a smart, semi-automated assistant that helps you find the bad tracks so you don't have to listen to all 64 of them. It focuses on two specific types of "noise" that are notoriously hard to find:

1. The Heartbeat Noise (EKG)

The Analogy: Imagine a drummer in the room who is beating a drum at a steady, rhythmic pace (the heartbeat). Sometimes, the sound of the drum leaks into the microphone and gets mixed into the brain conversation.

• How the tool works: The assistant doesn't just listen for a "thump." It looks for a specific musical pattern. A heartbeat isn't just one beat; it's a fundamental beat plus a series of harmonics (like a musical chord).
• The Magic: The tool scans the 64 tracks and asks, "Does this track have a rhythmic drum pattern with a specific musical structure?" If a track looks like it has that drum pattern, the assistant flags it and says, "Hey, check this one out."
• The Human Touch: The tool doesn't delete it automatically. It shows you the track's "map" (where the sound comes from on the head), the "waveform" (what the beat looks like), and the "spectrum" (the musical notes). You then click a button to say, "Yes, that's the drummer, delete it," or "No, that's actually part of the conversation, keep it."

2. The Digital "Beep" Noise (Trigger/DIN)

The Analogy: Imagine a computer in the room that sends a sharp, electronic "beep" every time a light flashes. Sometimes, that electrical "beep" leaks into the brain recording.

• The Problem: These beeps happen so fast and so regularly that they create a "comb" shape in the sound spectrum (like the teeth of a comb).
• How the tool works: The assistant knows exactly when the computer sends the beeps. It looks at the tracks and asks, "Does this track have energy exactly where the 'beep' teeth of the comb should be?"
• The Magic: If a track has that specific "comb" pattern, the assistant flags it.
• The Human Touch: Just like with the heartbeat, the tool shows you the visual evidence. You look at the "comb" pattern and decide if it's a glitch to be removed or a real signal.

Why This is a Big Deal

Before this tool, finding these specific noises was like looking for a needle in a haystack while wearing blindfold. You had to guess.

This tool is like giving you a metal detector.

1. It scans the whole haystack (all 64 tracks) in seconds.
2. It beeps only when it finds a needle (a likely artifact).
3. It hands you the needle and says, "Here are the top 3 suspects. You decide if you want to throw them away."

The Bottom Line

This paper isn't about replacing the human expert. It's about speeding up the search. It does the heavy lifting of sorting through the noise using math (spectral characteristics), but it keeps the human in the driver's seat for the final decision. It makes the process faster, more consistent, and much less frustrating for scientists trying to clean up their brain data.

Technical Summary

1. Problem Statement

Independent Component Analysis (ICA) is a standard preprocessing step for Electroencephalogram (EEG) data, used to decompose multichannel recordings into statistically independent components (ICs) to isolate neural signals from artifacts. However, a significant bottleneck remains in the manual identification and removal of artifact components.

• Challenges: Manual review is slow, subjective, and difficult to standardize.
• Specific Issues:
  • Cardiac (EKG) Artifacts: Often distributed across multiple components or hidden in lower-variance ICs, making them easy to miss if only high-variance components are inspected.
  • Trigger (Digital Input/DIN) Artifacts: Electrical leakage from digital event triggers (e.g., TTL pulses) can create sharp transients and harmonic structures that contaminate EEG channels.
• Gap: Existing tools (e.g., ICLabel, SASICA) are general-purpose. There is a need for targeted, semi-automated workflows that use specific spectral signatures to prioritize candidates for EKG and trigger leakage without removing human oversight.

2. Methodology

The authors developed the SENSI-EEG-Preproc-ICA-EKG-Trigger Module, a semi-automated MATLAB framework. The workflow operates post-ICA (after filtering, bad-channel handling, and ICA decomposition) and follows a "screen-then-review" logic:

A. Automated Detection Algorithms

The module uses frequency-domain metrics to score and rank ICs.

1. EKG Artifact Detection (autoDetectEkg):

   • Mechanism: Scans ICs for a physiologically plausible cardiac fundamental frequency ($f_0$, typically 0.8–2.0 Hz) and associated harmonic structure.
   • Metric: Calculates a Harmonic Local Signal-to-Noise Ratio (SNR).
     • It identifies a fundamental peak and defines narrow windows around expected harmonics ($h \cdot f_0$).
     • It computes the ratio of power in the harmonic peak windows against the local baseline (median power of surrounding frequencies).
     • A composite score is the sum of SNRs across multiple harmonics.
   • Ranking: Scores are standardized (robust Z-score using Median/MAD). Components with high outlier scores are flagged as strong candidates.

2. DIN (Trigger) Artifact Detection (autoDetectDIN):

   • Mechanism: Uses the known repetition period ($T_{DIN}$) of the digital trigger sequence to calculate the base frequency ($f_0 = 1/T_{DIN}$).
   • Metric: Calculates a Harmonic Energy Ratio.
     • It sums the power at expected harmonic frequencies ($h \cdot f_0$) within the analysis band.
     • It compares this to the median power of the surrounding baseline spectrum.
   • Ranking: The ratio is standardized across components. High positive Z-scores indicate energy concentrated specifically at trigger harmonics.

B. Interactive Review Interface

The automated scores do not automatically remove components. Instead, they populate a dedicated GUI (reviewEkgArtifactUI and reviewDINArtifactUI) where the user makes the final decision.

• Visualization: For each candidate, the interface displays three complementary views:
  1. Scalp Topography: Spatial distribution of the component.
  2. Time-Domain Activity: Representative segment of the IC time series.
  3. Frequency-Domain Structure: Power spectral density (PSD).
• User Control: Users toggle components between "Remove" (Red) and "Keep" (Green) based on the combined evidence.

3. Key Contributions

• Targeted Spectral Scoring: Unlike general classifiers, this module uses specific spectral rules (harmonic series for EKG; trigger-period harmonics for DIN) to detect artifacts that might be missed by variance-based sorting.
• Semi-Automated Workflow: It narrows the search space to a small set of high-probability candidates (typically top 3), significantly reducing manual review time while preserving interpretability.
• Open-Source Ecosystem: The release includes:
  • A public GitHub repository with the full MATLAB codebase.
  • A standalone example dataset (hosted on Stanford Digital Repository).
  • Comprehensive documentation and a user manual.
• Transparency: By retaining the human-in-the-loop for the final decision, the tool avoids the "black box" nature of fully automated classifiers, allowing researchers to document and justify artifact removal.

4. Results (Illustrative Analysis)

The paper demonstrates the workflow using an example dataset:

• EKG Detection: The algorithm successfully identified components with cardiac fundamentals and harmonics. In the review phase, the user confirmed that Component 9 (clear cardiac waveform) and Component 33 (slower cardiac contribution) were artifacts, while Component 48 (weaker evidence) was retained.
• DIN Detection: The algorithm identified a component (Component 113) with a "comb-like" harmonic spectrum matching the trigger frequency. The user confirmed this as an artifact, while other flagged candidates (Components 48 and 5) were retained as neural signals.
• Outcome: The tool successfully reduced the review burden by prioritizing specific components, allowing for rapid confirmation of artifacts that exhibit distinct spectral signatures.

5. Significance and Limitations

Significance:

• Reproducibility: Standardizes the identification of two common, difficult-to-detect artifact types.
• Efficiency: Drastically reduces the time required for post-ICA cleaning by focusing human attention only on likely candidates.
• Flexibility: Designed to be integrated into larger pipelines (e.g., the SENSI pipeline) and works with various EEG acquisition systems (specifically EGI/Net Station for DIN).

Limitations:

• Dependency on Preprocessing: Performance relies heavily on the quality of prior steps (filtering, bad-channel removal, and ICA decomposition). Poor ICA separation degrades the detector's accuracy.
• Specificity: The DIN detector requires a known trigger period and is only relevant for setups with digital trigger leakage.
• Parameter Sensitivity: Default parameters are tuned for the example data; users may need to adjust frequency bands or window sizes for different sampling rates or acquisition hardware.
• Spectral Overlap: There is a risk of removing neural responses (e.g., Steady-State Evoked Potentials) if they share harmonic frequencies with the trigger, though the review interface helps mitigate this.

Future Directions:
The authors suggest future work should focus on broader validation across diverse datasets and hardware, as well as exploring adaptive parameter selection to make the tool more robust to varying acquisition conditions.