ASR Validation Skill: SSVEP & Artifacts

Objective

Implement and validate Artifact Subspace Reconstruction (ASR) (specifically the clean_rawdata algorithm) using the 8-Channel SSVEP EEG Dataset with Artifact Trials (ds004745).

The core goal is to demonstrate that ASR can remove large-amplitude muscle and motion artifacts while preserving the underlying steady-state visually evoked potentials (SSVEP) at 2 Hz, 4 Hz, and 8 Hz.

Dataset Context

• Dataset ID: ds004745
• Features:
  • SSVEP Task: Users watch flickering lights at 2, 4, 8 Hz.
  • Artifact Trials: Users intentionally perform artifacts (head movement, jaw clench, eye blink) during the task.
  • Channels: 8 (Low channel count is a stress test for ASR).

Workflow Instructions

1. Data Ingestion

1. Download: Use datalad to download ds004745.

   datalad install https://github.com/OpenNeuroDatasets/ds004745.git
   cd ds004745
   datalad get sub-001/ses-01/eeg/*
   

2. Load: Load the raw EEG data (.set / .eeg). Note that some OpenNeuro datasets use EEGLAB format.
3. Locate Artifacts: Identify the trials/events marked as "Artifact" or "Movement". If explicit events are missing, inspect the time series for gross deviations (> 100 µV).

2. Implementation: JAX ASR

Implement the clean_rawdata algorithm in JAX using neurojax.preprocessing. If not present, you must implement:

1. Calibration: Calculate the geometric median and robust standard deviation (Huber/median absolute deviation) of the clean portions of the data (or a separate resting state file).
2. PCA Reconstruction:
   • Compute the sliding window PCA.
   • Identify components with variance > cutoff standard deviations (typically 5-20).
   • Reconstruct these high-variance components from the remaining subspace.

3. Verification & Validation (The "Test")

You must produce a validation_report.md containing:

1. Time-Domain Comparison:
   • Plot Raw vs ASR-Cleaned waveforms for an artifact segment.
   • Success Metric: Amplitude of artifact segments should be reduced to within physiological range (< 50-100 µV).
2. Frequency-Domain Comparison:
   • Compute the Power Spectral Density (PSD) for Raw vs ASR-Cleaned data.
   • Success Metric: The SSVEP peaks at 2 Hz, 4 Hz, and 8 Hz must remain visible and distinct in the cleaned data. They should NOT be attenuated significantly compared to non-artifact periods.
3. Quantitative Metric:
   • Signal-to-Noise Ratio (SNR): Calculate SNR of the SSVEP peaks before and after ASR.
   • Goal: SNR should increase (or at least decrease minimally) after cleaning.

Constraints

• Low Channel Count: The dataset has only 8 channels. ASR typically relies on spatial redundancy. You may need to tune the cutoff parameter (e.g., set it looser, around 20-30 SD) to avoid deleting real signal.
• JAX Compliance: All computations must be JAX-compatible for GPU acceleration.