Authors#

Please see the authors section in the CITATION.cff file for authors who have contributed to this software across releases.

Changelog#

Here we list what’s new in pyprep.

Version 0.7.1 (2026-06-15)#

Changelog#

  • Unified all median absolute deviation (MAD) computations in NoisyChannels behind a single direct-NumPy helper (pyprep.utils._mad), replacing the previous mix of scipy.stats.median_abs_deviation() calls and inline implementations. This keeps the speedup of the windowed loop in find_bad_by_correlation() while producing bit-identical output, by Stefan Appelhoff (#197)

  • NoisyChannels now detrends the EEG signal in place during initialization, avoiding two transient full-size copies of the recording; pyprep.removeTrend.removeTrend() gained a copy keyword argument (default True, preserving the previous behavior) to support this, by Stefan Appelhoff (#196)

  • find_bad_by_correlation() (and any other method that uses the internally band-pass-filtered signal) is now substantially faster: NoisyChannels._get_filtered_data filters all channels at once instead of channel-by-channel, with bit-identical output, by Stefan Appelhoff (#195)

Version 0.7.0 (2026-06-13)#

Changelog#

  • PrepPipeline can now be run in stages: the new remove_line_noise() and robust_reference() methods expose the individual PREP stages that fit() runs in sequence, and final bad-channel interpolation can be skipped via the new interpolate_bads argument of robust_reference() (and the matching pyprep.Reference.perform_reference() / pyprep.Reference.interpolate_bads() methods). These additions are fully backward compatible: fit() and all existing attributes are unchanged, by Austin Hurst and Stefan Appelhoff (#191)

  • find_bad_by_nan_flat() now respects the reject_by_annotation setting of the NoisyChannels instance, consistent with all other detection methods, by Roy Eric Wieske (#185)

  • Removed the unmaintained, non-asserting matplotlib-based PREP tests and the run_full_prep example (along with its committed MATLAB .mat artifacts); PyPREP’s fidelity to MATLAB PREP is validated by tests/test_matprep_compare.py instead, by Stefan Appelhoff (#163)

  • Added a new run_full_prep example demonstrating how to run the full pipeline on a noisy recording, inspect the detected bad channels, and compare the data before and after cleaning, by Stefan Appelhoff (#188)

  • The MATLAB PREP artifact-generation scripts (previously the external matprep_artifacts git submodule) are now vendored in tools/matprep/ and run by a new matprep_artifacts GitHub Actions workflow that publishes the reference artifacts used by tests/test_matprep_compare.py; the EEGLAB and MATLAB PREP versions are now pinned for reproducibility, by Stefan Appelhoff (#160)

Bug#

  • pyprep no longer calls logging.basicConfig() at import time (in pyprep.reference), which previously reconfigured the root logger of any downstream application that imported pyprep, by Stefan Appelhoff

  • Reference no longer emits a spurious “No bad channels to interpolate” warning during robust referencing when there are no bad channels to interpolate, by Stefan Appelhoff (#192)

Code health#

  • Added support for Python 3.14 to the test matrix and package metadata, and bumped the pre-commit hooks and GitHub Actions used for development and CI, by Stefan Appelhoff (#192)

Version 0.6.0 (2026-02-26)#

Changelog#

Version 0.5.0 (2025-07-17)#

Changelog#

Bug#

Code health#

Version 0.4.3 (2023-10-27)#

Changelog#

Bug#

Version 0.4.2 (2022-03-16)#

Changelog#

  • Fixing an issue in .zenodo.json metadata from 0.4.1.

Version 0.4.1 (2022-03-16)#

Changelog#

  • This is a patch release, including a short documentation update (#110) and several infrastructure fixes. Most importantly, this version is released to trigger an archive at Zenodo via the GitHub integration.

Version 0.4.0 (2021-10-22)#

Changelog#

  • Created a new module named pyprep.ransac which contains find_bad_by_ransac, a standalone function mirroring the previous ransac method from the NoisyChannels class, by Yorguin Mantilla (#51)

  • Added two attributes PrepPipeline.noisy_channels_before_interpolation and PrepPipeline.noisy_channels_after_interpolation which have the detailed output of each noisy criteria, by Yorguin Mantilla (#45)

  • Added two keys to the PrepPipeline.noisy_channels_original dictionary: bad_by_dropout and bad_by_SNR, by Yorguin Mantilla (#45)

  • Changed RANSAC chunking logic to reduce max memory use and prefer equal chunk sizes where possible, by Austin Hurst (#44)

  • Changed RANSAC’s random channel sampling code to produce the same results as MATLAB PREP for the same random seed, additionally changing the default RANSAC sample size from 25% of all good channels (e.g. 15 for a 64-channel dataset with 4 bad channels) to 25% of all channels (e.g. 16 for the same dataset), by Austin Hurst (#62)

  • Changed RANSAC so that “bad by high-frequency noise” channels are retained when making channel predictions (provided they aren’t flagged as bad by any other metric), matching MATLAB PREP behaviour, by Austin Hurst (#64)

  • Added a new flag matlab_strict to PrepPipeline, Reference, NoisyChannels, and find_bad_by_ransac() for optionally matching MATLAB PREP’s internal math as closely as possible, overriding areas where PyPREP attempts to improve on the original, by Austin Hurst (#70)

  • Added a matlab_strict method for high-pass trend removal, exactly matching MATLAB PREP’s values if matlab_strict is enabled, by Austin Hurst (#71)

  • Added a window-wise implementation of RANSAC and made it the default method, reducing the typical RAM demands of robust re-referencing considerably, by Austin Hurst (#66)

  • Added max_chunk_size parameter for specifying the maximum chunk size to use for channel-wise RANSAC, allowing more control over PyPREP RAM usage, by Austin Hurst (#66)

  • Changed Reference to exclude “bad-by-SNR” channels from initial average referencing, matching MATLAB PREP behaviour, by Austin Hurst (#78)

  • Changed Reference to only flag “unusable” channels (bad by flat, NaNs, or low SNR) from the first pass of noisy detection for permanent exclusion from the reference signal, matching MATLAB PREP behaviour, by Austin Hurst (#78)

  • Added a framework for automated testing of PyPREP’s components against their MATLAB PREP counterparts (using .mat and .set files generated with the matprep_artifacts script), helping verify that the two PREP implementations are numerically equivalent when matlab_strict is True, by Austin Hurst (#79)

  • Changed NoisyChannels to reuse the same random state for each run of RANSAC when matlab_strict is True, matching MATLAB PREP behaviour, by Austin Hurst (#89)

  • Added a new argument as_dict for get_bads(), allowing easier retrieval of flagged noisy channels by category, by Austin Hurst (#93)

  • Added a new argument max_iterations for perform_reference() and robust_reference(), allowing the maximum number of referencing iterations to be user-configurable, by Austin Hurst (#93)

  • Changed robust_reference() to ignore bad-by-dropout channels during referencing if matlab_strict is True, matching MATLAB PREP behaviour, by Austin Hurst (#93)

  • Changed robust_reference() to allow initial bad-by-SNR channels to be used for rereferencing interpolation if no longer bad following initial average reference, matching MATLAB PREP behaviour, by Austin Hurst (#93)

  • Added a matlab_strict method for bad channel interpolation, allowing for full numeric equivalence with MATLAB PREP’s robust referencing, by Austin Hurst (#96)

Bug#

  • Fixed RANSAC to give consistent results with a fixed seed across different chunk sizes, by Austin Hurst and Yorguin Mantilla (#43)

  • Fixed “bad channel by flat” threshold in NoisyChannels.find_bad_by_nan_flat() to be consistent with MATLAB PREP, by Austin Hurst (#60)

  • Changed “bad channel by deviation” and “bad channel by correlation” detection code in NoisyChannels to compute IQR and quantiles in the same manner as MATLAB, thus producing identical results to MATLAB PREP, by Austin Hurst (#57)

  • Fixed a bug where EEG data was getting reshaped into RANSAC windows incorrectly (channel samples were not sequential), which was causing considerable variability and noise in RANSAC results, by Austin Hurst (#67)

  • Fixed RANSAC to avoid making unnecessary signal predictions for known-bad channels, matching MATLAB behaviour and reducing RAM requirements, by Austin Hurst (#72)

  • Fixed a bug in NoisyChannels.find_bad_by_correlation() that prevented it from being able to handle channels with dropouts (intermittent flat regions), by Austin Hurst (#81).

  • Fixed NoisyChannels so that it always runs “bad channel by NaN” and “bad channel by flat” detection, preventing these channels from causing problems with other NoisyChannels methods, by Austin Hurst (#79)

  • Fixed Reference so that channels are no longer excluded from final average reference calculation if they were originally bad by NaN, flat, or low SNR, by Austin Hurst (#92)

API#

  • The permissible parameters for the following methods were removed and/or reordered: ransac._ransac_correlations, ransac._run_ransac, and ransac._get_ransac_pred methods, by Yorguin Mantilla (#51)

  • The following methods have been moved to a new module named ransac and are now private: NoisyChannels.ransac_correlations, NoisyChannels.run_ransac, and NoisyChannels.get_ransac_pred methods, by Yorguin Mantilla (#51)

  • The permissible parameters for the following methods were removed and/or reordered: NoisyChannels.ransac_correlations, NoisyChannels.run_ransac, and NoisyChannels.get_ransac_pred methods, by Austin Hurst and Yorguin Mantilla (#43)

  • Changed the meaning of the argument channel_wise in find_bad_by_ransac() to mean ‘perform RANSAC across chunks of channels instead of window-wise’, from its original meaning of ‘perform channel-wise RANSAC one channel at a time’, by Austin Hurst (#66)

  • The arguments fraction_bad and fraction_good were renamed to frac_bad and sample_prop, respectively, for find_bad_by_ransac() and find_bad_by_ransac(), by Austin Hurst (#88)

  • The units of all user-facing EEG data attributes in PrepPipeline and Reference were changed from mV to Volts to better match MNE convention, by Austin Hurst (#102)

Version 0.3.1 (2020-10-10)#

Changelog#

  • It’s now possible to pass keyword arguments to the notch filter inside PrepPipeline; see the filter_kwargs parameter by Yorguin Mantilla (#40)

  • The default filter length for the spectrum_fit method will be ’10s’ to fix memory issues, by Yorguin Mantilla (#40)

  • Channel types are now available from a new ch_types_all attribute, and non-EEG channel names are now available from a new ch_names_non_eeg attribute from PrepPipeline, by Yorguin Mantilla (#34)

  • Renaming of ch_names attribute of PrepPipeline to ch_names_all, by Yorguin Mantilla (#34)

  • It’s now possible to pass 'eeg' to ref_chs and reref_chs keywords to the prep_params parameter of PrepPipeline to select only eeg channels for referencing, by Yorguin Mantilla (#34)

  • PrepPipeline will retain the non eeg channels through the raw attribute. The eeg-only and non-eeg parts will be in raw_eeg and raw_non_eeg respectively. See the raw attribute, by Christian O’Reilly (#34)

  • When a ransac call needs more memory than available, pyprep will now automatically switch to a slower but less memory-consuming version of ransac, by Yorguin Mantilla (#32)

  • It’s now possible to pass an empty list for the line_freqs param in PrepPipeline to skip the line noise removal, by Yorguin Mantilla (#29)

  • The three main classes PrepPipeline, NoisyChannels, and pyprep.Reference now have a random_state parameter to set a seed that gets passed on to all their internal methods and class calls, by Stefan Appelhoff (#31)

Bug#

API#

  • Remove noisy.py module from the pyprep package. Its main functionality has been migrated to the remaining modules, and the functions for FASTER have been dropped because they were out of scope, by Stefan Appelhoff (#39)

Version 0.3.0 (2020-01-04)#

Changelog#

  • Include a boolean do_detrend in robust_reference() to indicate whether detrend should be done internally or not for the use with NoisyChannels, by Yorguin Mantilla (#9)

  • Robust average referencing + tests, by Victor Xiang (#6)

  • Removing trend in the EEG data by high pass filtering and local linear regression + tests, by Aamna Lawrence (#6)

  • Finding noisy channels with comparable output to Matlab + tests-including test for ransac, by Aamna Lawrence (#6)

  • Stringing all the things together for the PREP pipeline + tests, by Victor Xiang (#6)

  • Finding noisy channels with comparable output to Matlab + tests-including test for ransac, by Aamna Lawrence (#6)

  • Finding the appropriate parameters in the MNE notch filter for implementing clean line noise functionality of Eeglab, by Aamna Lawrence (#6)

  • Finding the reason for the difference between the Matlab and Pyprep’s output- Probably minor differences in the filter functions and also rounding done by functions like quantile, by Victor Xiang and Aamna Lawrence (#6)

Bug#

API#

  • Oversaw modularization of PREP pipeline’s submodules and a scikit learn style pyprep.fit, by Adam Li (#6)

  • Oversaw ChangeLog by Victor Xiang and Aamna Lawrence for transitioning pyprep to Version 0.3.0, by Adam Li (#6)

Version 0.2.3 (2019-02-02)#

Doc#

  • update formatting and docs and fix tags and releases post-hoc, by Stefan Appelhoff

Version 0.2.2 (2018-05-01)#

Bug#

Version 0.2.1 (2018-04-29)#

Changelog#

Version 0.1.1 (2018-04-29)#

Changelog#

  • Miscellaneous changes

Version 0.1.0 (2018-04-24)#

Changelog#

  • Initial commit: 2018-04-12

  • Miscellaneous changes