Filtering#

This section will guide you through the process of filtering EEG data using MNE-Python. Filtering is a crucial step in EEG data preprocessing, as it helps to remove noise and artifacts from the signal.

Prerequisites

This guide assumes you have loaded your EEG data into a raw object. See Load Data for details on loading EDF files.

Understanding Filtering#

Filters remove unwanted frequency components from your signal:

  • High-pass filter: Removes slow drifts (< cutoff frequency)

  • Low-pass filter: Removes high-frequency noise (> cutoff frequency)

  • Band-pass filter: Keeps only frequencies within a range

  • Notch filter: Removes specific frequencies (e.g., 50/60 Hz line noise)

Filtering Resources

See Background on Filtering and Filtering and Resampling for comprehensive filtering guidance.

Basic Filtering#

Band-Pass Filter#

The most common filter type - keeps frequencies within a specified range:

Apply band-pass filter to remove drift and high-frequency noise#
import mne

# Load data
raw = mne.io.read_raw_edf('recording.edf', preload=True)

# Visualize before filtering
raw.plot(n_channels=20, duration=10, title='Before Filtering')

# Apply band-pass filter (1-40 Hz)
raw.filter(l_freq=1.0, h_freq=40.0)

# Visualize after filtering
raw.plot(n_channels=20, duration=10, title='After Filtering')

Common frequency ranges:

  • ERP analysis: 0.1 - 50 Hz

  • Spectral analysis: 0.5 - 50 Hz

  • Beta/gamma analysis: 1 - 100 Hz

  • Sleep staging: 0.3 - 35 Hz

High-Pass Filter Only#

Remove slow drifts and DC offset:

Remove slow drifts below 1 Hz#
# Remove frequencies below 1 Hz
raw.filter(l_freq=1.0, h_freq=None)

Low-Pass Filter Only#

Remove high-frequency noise:

Remove high-frequency noise above 40 Hz#
# Remove frequencies above 40 Hz
raw.filter(l_freq=None, h_freq=40.0)

Notch Filtering#

Remove power line noise (50 Hz in Europe, 60 Hz in North America):

Remove 60 Hz line noise and harmonics#
# Remove 60 Hz line noise
raw.notch_filter(freqs=60.0, filter_length='auto', phase='zero')

# Remove 60 Hz and harmonics
raw.notch_filter(freqs=[60, 120, 180], filter_length='auto', phase='zero')

# For 50 Hz regions
raw.notch_filter(freqs=50.0, filter_length='auto', phase='zero')

Visualizing Line Noise#

Check if line noise is present before filtering:

Plot power spectral density to identify line noise#
# Plot power spectral density
raw.plot_psd(fmax=100, average=False)
# Look for peaks at 50/60 Hz and harmonics

Filter Parameters#

Filter Type (FIR vs IIR)#

FIR (Finite Impulse Response) - Default:

  • Linear phase (no temporal distortion)

  • Symmetric impulse response

  • Longer computational time

  • Recommended for most EEG applications

Apply FIR filter#
# FIR filter (default)
raw.filter(l_freq=1.0, h_freq=40.0, method='fir')

IIR (Infinite Impulse Response):

  • Non-linear phase (can distort timing)

  • Sharper frequency cutoffs

  • Faster computation

  • Use with caution for ERP analysis

Apply IIR Butterworth filter#
# IIR filter (4th order Butterworth)
raw.filter(l_freq=1.0, h_freq=40.0, method='iir', iir_params=dict(order=4, ftype='butter'))

Filter Window#

Controls the shape of the filter frequency response:

Choose filter window type#
# Hamming window (default)
raw.filter(l_freq=1.0, h_freq=40.0, fir_window='hamming')

# Hann window (smoother frequency response)
raw.filter(l_freq=1.0, h_freq=40.0, fir_window='hann')

# Blackman window (better stopband attenuation)
raw.filter(l_freq=1.0, h_freq=40.0, fir_window='blackman')

Filter Length#

Longer filters have sharper cutoffs but introduce more edge artifacts:

Specify filter length#
# Automatic length (recommended)
raw.filter(l_freq=1.0, h_freq=40.0, filter_length='auto')

# Specific length (in samples)
raw.filter(l_freq=1.0, h_freq=40.0, filter_length='10s')  # 10 seconds

# Very sharp filter
raw.filter(l_freq=1.0, h_freq=40.0, filter_length='30s')

Phase#

Controls whether filtering introduces time delays:

Choose filter phase behavior#
# Zero-phase (default) - no time delay, but processes data twice
raw.filter(l_freq=1.0, h_freq=40.0, phase='zero')

# Minimum phase - introduces time delay but processes once
raw.filter(l_freq=1.0, h_freq=40.0, phase='minimum')

Advanced Filtering#

Transition Bandwidth#

Control the sharpness of frequency cutoffs:

Set transition bandwidth for cutoff sharpness#
# Wider transition (faster, less ringing)
raw.filter(l_freq=1.0, h_freq=40.0, l_trans_bandwidth=0.5, h_trans_bandwidth=5.0)

# Narrower transition (sharper cutoff, more ringing)
raw.filter(l_freq=1.0, h_freq=40.0, l_trans_bandwidth=0.1, h_trans_bandwidth=1.0)

# Automatic (recommended)
raw.filter(l_freq=1.0, h_freq=40.0, l_trans_bandwidth='auto', h_trans_bandwidth='auto')

Filtering Specific Channels#

Filter only selected channels#
# Filter only EEG channels
raw.filter(l_freq=1.0, h_freq=40.0, picks='eeg')

# Filter specific channels
raw.filter(l_freq=1.0, h_freq=40.0, picks=['EEG Cz-LE', 'EEG Pz-LE'])

Filtering in Place vs Copy#

Filter in-place or create filtered copy#
# Modify original data (in-place)
raw.filter(l_freq=1.0, h_freq=40.0)

# Create filtered copy (preserves original)
raw_filtered = raw.copy().filter(l_freq=1.0, h_freq=40.0)

Filter Design Visualization#

Visualize filter characteristics before applying:

Visualize filter frequency and impulse response#
from mne.viz import plot_filter

# Design filter parameters
h = mne.filter.create_filter(
    raw.get_data(), raw.info['sfreq'],
    l_freq=1.0, h_freq=40.0,
    method='fir', fir_window='hamming'
)

# Plot frequency and impulse response
plot_filter(h, raw.info['sfreq'])

Resampling#

Reduce sampling rate to decrease file size and processing time:

Resample data to lower sampling rate#
# Original sampling rate
print(f'Original sampling rate: {raw.info["sfreq"]} Hz')

# Resample to 150 Hz
raw.resample(sfreq=150)

print(f'New sampling rate: {raw.info["sfreq"]} Hz')

Filter Before Resampling

Always filter BEFORE resampling to avoid aliasing artifacts. The correct order is:

  1. Notch filter (if needed)

  2. Band-pass filter

  3. Resample

Comparing Filter Effects#

Visualize the impact of different filters:

Compare filter effects using power spectral density#
# Create copies with different filters
raw_1hz = raw.copy().filter(l_freq=1.0, h_freq=40.0)
raw_01hz = raw.copy().filter(l_freq=0.1, h_freq=40.0)

# Plot power spectral density for comparison
import matplotlib.pyplot as plt

fig, axes = plt.subplots(3, 1, figsize=(10, 8))

raw.plot_psd(ax=axes[0], show=False, average=True)
axes[0].set_title('Unfiltered')

raw_1hz.plot_psd(ax=axes[1], show=False, average=True)
axes[1].set_title('1-40 Hz')

raw_01hz.plot_psd(ax=axes[2], show=False, average=True)
axes[2].set_title('0.1-40 Hz')

plt.tight_layout()
plt.show()

Next Steps#

After filtering your data:

  1. Remove artifacts with ICA - Clean data for analysis

  2. Create epochs - Extract event-related segments

  3. MNE-LSL real-time filtering - Compare with real-time approaches

Resources#