Artifact Subspace
Reconstruction (ASR)

Imagine you are recording someone wearing an EEG headset while walking or exercising. Every time they bite down, move their jaw, or twitch a neck muscle, a massive electrical spike blasts across the sensors.

Traditional cleaning gives you two bad choices:

• Throw away the whole 1-second segment: If you do this every time the subject twitches, you lose over 50% of your recording!
• Run ICA (Independent Component Analysis): ICA works great for steady, repeating eye blinks. But for random, sudden muscle pops, ICA gets confused because the muscle pop doesn't repeat consistently across the recording.

The Real-World Analogy: The Building Inspector

🤝 Why ASR is the Essential Pre-conditioner for ICA (Stationarity vs. Bursts)

Students often ask: "If we have ICA, why do we need ASR?" The answer lies in the mathematical assumptions of both methods:

Before ASR can clean artifacts, it needs to know what clean, baseline brain activity looks like for this specific person and electrode layout.

We collect a short calibration recording (e.g. 1 minute resting state with eyes open/closed, or automatically find the cleanest 15% segments of the dataset).

Step 2 — The 500 ms Sliding Spotlight: As the continuous EEG recording streams, ASR slides a short 500 ms window across the data sample by sample. Inside this window, Principal Component Analysis (PCA) decomposes the multi-channel signal into independent directional axes ($v_1, v_2, \dots, v_Q$).

Step 3 — The Cutoff Rule ($k$ Standard Deviations): For each component direction, ASR asks: "Is the variance in this direction abnormally huge compared to baseline?"

📊 Empirical Cutoff ($k$) Guidelines from Literature

Early studies suggested $k = 5$ for real-time BCI, but landmark validation studies (Chang et al., 2018; Blum et al., 2019) proved that low $k$ values over-clean and attenuate genuine Event-Related Potentials (ERPs). Follow these evidence-based thresholds:

Here comes the magical step of ASR: How do we repair the corrupted signal without dropping the data window?

When component 1 is flagged as an artifact, ASR zeroes out component 1 in diagonal matrix $U$ ($U_{11}=0$). Then, it uses the Reference Mixing Matrix $M$ (which knows how healthy brain channels normally talk to each other) to predict what brain wave should have been there!

Production ASR Pipeline & Code Reference

import mne
from meegkit.asr import ASR

# 1. Load continuous raw EEG dataset
raw = mne.io.read_raw_fif("sample_raw.fif", preload=True)

# 2. Mandatory: High-pass filter at 1.0 Hz (ASR requires zero-mean data)
# DO NOT apply Common Average Referencing (CAR) before ASR!
raw.filter(l_freq=1.0, h_freq=40.0)

# 3. Pick EEG channels and extract numpy array
raw_eeg = raw.copy().pick('eeg')
data = raw_eeg.get_data()
sfreq = raw.info['sfreq']

# 4. Initialize ASR with empirical research standard cutoff (k = 20 to 30)
asr = ASR(sfreq=sfreq, cutoff=20)

# 5. Fit ASR on 60 seconds of clean resting baseline (or cleanest segment)
calibration_samples = int(60 * sfreq)
asr.fit(data[:, :calibration_samples])

# 6. Apply subspace reconstruction across continuous EEG
clean_data, _ = asr.transform(data)
raw_eeg._data = clean_data

# 7. NOW apply Common Average Reference (CAR) and proceed to ICA
raw_eeg.set_eeg_reference('average')
print("ASR Cleaning Complete! Ready for stationary ICA.")

% 1. Load continuous raw dataset into EEGLAB
EEG = pop_loadset('filename', 'sample_raw.set');

% 2. Run clean_rawdata pipeline with standard research parameters
% Note: clean_artifacts handles highpass, bad channels, ASR, and CAR automatically
EEG_clean = clean_artifacts(EEG, ...
    'FlatlineCriterion', 5, ...       % Drop channels flatlining > 5s
    'Highpass', [0.5 1.0], ...         % Transition band for baseline removal
    'ChannelCriterion', 0.8, ...       % Remove channels with low correlation < 0.8
    'LineNoiseCriterion', 4, ...       % Drop channels with extreme 50/60 Hz noise
    'BurstCriterion', 20, ...          % ASR Cutoff (k = 20: Recommended Research Standard)
    'WindowCriterion', 0.25, ...       % Drop epochs with > 66% corrupted channels
    'BurstRejection', 'on', ...        % Enable ASR subspace reconstruction
    'Distance', 'Euclidian');

% 3. Apply Common Average Reference (CAR) AFTER ASR
EEG_clean = pop_reref(EEG_clean, []);

% 4. Proceed to ICA with clean stationary data
EEG_clean = pop_runica(EEG_clean, 'extended', 1);

Student Self-Check Quiz