Principal Component Analysis - Optimal Basis Sets (PCA-OBS) removing cardiac artefact#

This script shows an example of how to use an adaptation of PCA-OBS [1]. PCA-OBS was originally designed to remove the ballistocardiographic artefact in simultaneous EEG-fMRI. Here, it has been adapted to remove the delay between the detected R-peak and the ballistocardiographic artefact such that the algorithm can be applied to remove the cardiac artefact in EEG (electroencephalography) and ESG (electrospinography) data. We will illustrate how it works by applying the algorithm to ESG data, where the effect of removal is most pronounced.

See: https://www.biorxiv.org/content/10.1101/2024.09.05.611423v1 for more details on the dataset and application for ESG data.

# Authors: Emma Bailey <bailey@cbs.mpg.de>,
#          Steinn Hauser Magnusson <hausersteinn@gmail.com>
# License: BSD-3-Clause
# Copyright the MNE-Python contributors.

import glob

import numpy as np

Download sample subject data from OpenNeuro if you haven’t already. This will download simultaneous EEG and ESG data from a single run of a single participant after median nerve stimulation of the left wrist.

import openneuro
from matplotlib import pyplot as plt

import mne
from mne import Epochs, events_from_annotations
from mne.io import read_raw_eeglab
from mne.preprocessing import find_ecg_events, fix_stim_artifact

# add the path where you want the OpenNeuro data downloaded. Each run is ~2GB of data
ds = "ds004388"
target_dir = mne.datasets.default_path() / ds
run_name = "sub-001/eeg/*median_run-03_eeg*.set"
if not glob.glob(str(target_dir / run_name)):
    target_dir.mkdir(exist_ok=True)
    openneuro.download(dataset=ds, target_dir=target_dir, include=run_name[:-4])
block_files = glob.glob(str(target_dir / run_name))
assert len(block_files) == 1

Define the esg channels (arranged in two patches over the neck and lower back).

esg_chans = [
    "S35",
    "S24",
    "S36",
    "Iz",
    "S17",
    "S15",
    "S32",
    "S22",
    "S19",
    "S26",
    "S28",
    "S9",
    "S13",
    "S11",
    "S7",
    "SC1",
    "S4",
    "S18",
    "S8",
    "S31",
    "SC6",
    "S12",
    "S16",
    "S5",
    "S30",
    "S20",
    "S34",
    "S21",
    "S25",
    "L1",
    "S29",
    "S14",
    "S33",
    "S3",
    "L4",
    "S6",
    "S23",
]

# Interpolation window in seconds for ESG data to remove stimulation artefact
tstart_esg = -7e-3
tmax_esg = 7e-3

# Define timing of heartbeat epochs in seconds relative to R-peaks
iv_baseline = [-400e-3, -300e-3]
iv_epoch = [-400e-3, 600e-3]

Next, we perform minimal preprocessing including removing the stimulation artefact, downsampling and filtering.

raw = read_raw_eeglab(block_files[0], verbose="error")
raw.set_channel_types(dict(ECG="ecg"))
# Isolate the ESG channels (include the ECG channel for R-peak detection)
raw.pick(esg_chans + ["ECG"])
# Trim duration and downsample (from 10kHz) to improve example speed
raw.crop(0, 60).load_data().resample(2000)

# Find trigger timings to remove the stimulation artefact
events, event_dict = events_from_annotations(raw)
trigger_name = "Median - Stimulation"

fix_stim_artifact(
    raw,
    events=events,
    event_id=event_dict[trigger_name],
    tmin=tstart_esg,
    tmax=tmax_esg,
    mode="linear",
    stim_channel=None,
)

Find ECG events and add to the raw structure as event annotations.

ecg_events, ch_ecg, average_pulse = find_ecg_events(raw, ch_name="ECG")
ecg_event_samples = np.asarray(
    [[ecg_event[0] for ecg_event in ecg_events]]
)  # Samples only

qrs_event_time = [
    x / raw.info["sfreq"] for x in ecg_event_samples.reshape(-1)
]  # Divide by sampling rate to make times
duration = np.repeat(0.0, len(ecg_event_samples))
description = ["qrs"] * len(ecg_event_samples)

raw.annotations.append(
    qrs_event_time, duration, description, ch_names=[esg_chans] * len(qrs_event_time)
)

Create evoked response around the detected R-peaks before and after cardiac artefact correction.

events, event_ids = events_from_annotations(raw)
event_id_dict = {key: value for key, value in event_ids.items() if key == "qrs"}
epochs = Epochs(
    raw,
    events,
    event_id=event_id_dict,
    tmin=iv_epoch[0],
    tmax=iv_epoch[1],
    baseline=tuple(iv_baseline),
)
evoked_before = epochs.average()

# Apply function - modifies the data in place. Optionally high-pass filter
# the data before applying PCA-OBS to remove low frequency drifts
raw = mne.preprocessing.apply_pca_obs(
    raw, picks=esg_chans, n_jobs=5, qrs_times=raw.times[ecg_event_samples.reshape(-1)]
)

epochs = Epochs(
    raw,
    events,
    event_id=event_id_dict,
    tmin=iv_epoch[0],
    tmax=iv_epoch[1],
    baseline=tuple(iv_baseline),
)
evoked_after = epochs.average()

Compare evoked responses to assess completeness of artefact removal.

fig, axes = plt.subplots(1, 1, layout="constrained")
data_before = evoked_before.get_data(units=dict(eeg="uV")).T
data_after = evoked_after.get_data(units=dict(eeg="uV")).T
hs = list()
hs.append(axes.plot(epochs.times, data_before, color="k")[0])
hs.append(axes.plot(epochs.times, data_after, color="green", label="after")[0])
axes.set(ylim=[-500, 1000], ylabel="Amplitude (µV)", xlabel="Time (s)")
axes.set(title="ECG artefact removal using PCA-OBS")
axes.legend(hs, ["before", "after"])
plt.show()

References#

Estimated memory usage: 0 MB

Gallery generated by Sphinx-Gallery