v1.1.0

[1.1.0] - 2026-02-21

Changed

• Added compatibility support for Python 3.14 and pandas 3.0.

Fixed

• Fixed an edge case in refine_gs where gait sequences with zero detected initial contacts could raise an error.
  (#227)
• Updated the Mobilise-D pipeline per-GS loop to skip GSs with fewer than two initial contacts, avoiding crashes while
  still producing consistent empty outputs when no valid GSs remain.
  (#227)

Added

• Added proxy values for all intermediary pipeline outputs the to MobilisedPipelineUniversal class.
  This allows for easier access to the intermediary results of underlying pipeline.

v1.0.0 - 🎉🎉

[1.0.0] - 2025-07-11

This marks the official first stable release of the mobgap package.
This means that all core algorithms are stable and have been validated against the old Matlab implementations.
To learn more about this release, check out our short blogpost on the release:
https://mobgap.readthedocs.io/en/latest/blog/20250708_1_0.html

Changed

• When loading reference data using the Mobilise-D Matlab format, the handling of duplicated ICs and LR labels has been
  updated.
  Namely, we moved the processing of duplicated ICs to a WB level. This means duplicated ICs across WB are not removed
  or deduplicated anymore.
  Further, we actually drop duplicated ICs also from the IC array.
  We decided that this is the best behavior given how this data is usually used.
  The reason for that is, that they are cases where, due to the way the INDIP processes the data, WBs exist where one
  WB ends with the same IC the next WB starts, basically creating WBs without break between them.
  While we generally consider this a bug in the INDIP processing, it is something that is out of scope to fix in
  existing data.
  Hence, we are excepting this as a plausible case and don't fix it.
• Fixed multiple edgecases where unexpected errors were thrown when data sequences were too short or empty.
  Algorithms now correctly return empty results instead of throwing an error.
• Introduced a calculation of raw number of initial contacts per WB into the WBA. This parameter is now calculated by
  default in all pipelines and also aggregated by the MobiliseDAggregator.
• The aggregated parameters of MobiliseDAggregator are renamed based on new Mobilise-D internal consensus:
  • walkdur_all_sum is now provided in minutes instead of hours. The alternative name is changed to
    total_walking_duration_min
  • *_max are all renamed to *_p90 to reflect that they are the 90th percentile of the values and not really the
    maximum.
  • steps_all_sum is renamed to wbsteps_all_sum to reflect that it is not the sum of all steps, but only considers
    initial contacts within valid WBs. The new preferred alternative name is renamed to
    wb_all__n_raw_initial_contacts__sum to signal even further that we are counting raw ICs and not really steps.

v0.11.0 - Getting ready for 1.0

[0.11.0] - 2025-06-16

This is the last release before 1.0.
Most things are stable now and the algorithms published in this release have been validated against there old
implementations.
Results of this validation can be in the docs under the new "Revalidation" section.
Raw results for the revalidation can be found in the companion
revalidation repository.

Beyond that this release contains the last expected set of changes to the main pipelines.
With the 1.0 release, we will expect the outputs of the main pipelines to be stable (unless actual bugs are found).
Potential improvements to algorithms will always require manual opt-in.

SCIENTIFIC CHANGES

• The default ICC calculation was changed from ICC1,1 to ICC2,1
  (#176)
• The GsdIluz algorithm was reworked. This fixes some discrepancies with the original implementation and should improve
  the results in many cases.
  We now also provide a version of the original peak detection algorithm that was used in the matlab implementation.
  This can be used by setting `GsdIluz(**GsdIluz.PredefinedParameters.original).
  (#182, #187)
• The models behind the Left-Right detection algorithm by Ullrich et al. have been retrained and made reproducible.
  This results in some changes in the results (mostly improvements) and some changes in the output, even for the old
  models, as parts of the pre-processing was adapted to better reflect the original implementation.
  (#186)

Added

• We added first scripts for the revalidation.
  This includes a script to extract the old results and put them into files the new evaluation pipeline can read.
  Further, we have a first sceleton for what the gsd-revalidation could look like.
  (#182)
• All pipelines now pass much more metadata to the algorithms action methods.
  This should make it easier to implement algorithms that need more context.
  At the moment this is only used in the DummyAlgorithm to load existing results based on the participant and test.
  For this to work across all algorithms, we had to change some logic in the LrcUllrich algorithm to selectively
  parse the kwargs that it needs.
  (#182)
• When using Custom Aggregations, it is now easier to pass arguments to the underlying aggregation functions.
  This is a breaking change, as we require you to now pass the desired name of the output column as part of the
  column_name.
  CustomOperation(... func=A.icc, column_name="my_column") -> CustomOperation(... func=A.icc, column_name=("icc", "my_column"))
  This removes the requirement that func needs to be a proper function object with a __name__ attribute.
  Hence, you can now use lambdas and partials without issue.
  (#176)
• It is now possible to split custom aggregations into multiple columns.
  For example, A.icc returned two values, the ICC and the confidence intervals.
  Before, both values where put into the same column.
  Now you can specify two column names in Custom Aggregation to split the values.
  CustomOperation(... func=A.icc, column_name=[("icc", "my_column"), ("icc_ci", "my_ci_column")])
  (#176)
• Confidence intervals where added as a aggregation function.
  (#176)
• Helper for styling and formatting tables. They are primarily used within the revalidation docs.
  But feel free to copy them for your own use.
  (#192 , #199)
• Plotting helper for validation plots (#204)
• The MSDataset now has a dedicated dataset class that can be used to work with it. The dataset will be published on
  Zenodo in the future. (#186)

Fixed

• Obscure bug when loading the TVS dataset causing a IntCastingNaNError, when the participant_information.xlsx file
  was opened and resaved using Microsoft Excel. For more infos see (#197)
• When calculating the total wear time during waking hours, waking hours were not considered correctly and were affected
  by the daylight saving state at the time the script was run! This is now fixed, by not using actual dates for the
  comparison, but just seconds from midnight.
• apply_transformation now returns a normal index instead of MultiIndex, when all expected output col names are
  strings (#199)

Removed

• In anticipation of the tpcp>2.0 update, we removed the score methods from all pipelines.
  You now need to provide the scorer directly to the MetaPipelines/Optimizers.
  (#182)

v0.10.0 - Speed improvements and QoL for TVS dataset

[0.10.0] - 2024-10-22

Added

• Added a simple way to add performance (as in "time algo needs to run") tracking to algorithms
• Performance values are now reported by the GSD algorithms and the full pipelines via the perf_ attribute.
• The TVS datasets now have two new columns {test/recording}_name and {test/recording}_name_pretty that have easier
  to read names for the tests and recordings.

Changed

• Swapped out the peak detection per window in GSD Iluz to a custom vectorized one that can be jit compiled.
  This provides a 2-3x speedup for large inputs.
• Critical path in GSD-Ionesco is now jit compiled.

For Developers

• We now support adding a .env file to the root of the project to set environment variables and have them loaded
  automatically when running the tests or examples for the TVS dataset.
  When test depend on env vars, use the mobgap.misc.get_env_var function to access them.

v0.9.0 - Fixed official aggregation script + some fixes

[0.9.0] - 2024-10-21

Changed

• The calculate_matched_gsd_performance_metrics function now always return the error metrics that depend on TN
  samples, not just when TN samples exist.
  This way the output structure is consistent, and we can avoid bugs in scorer functions, where some datapoints might
  unexpectedly return a different set of error metrics.
• The official script to aggregate DMOs, now correctly converts stride length values to cm and variance values to cm^2.
  Note, that the internal function within mobgap, does not do this conversion automatically, as we use different units
  in mobgap internally. If you need to match the official script, you have to do this conversion manually.
  see examples/aggregation/_99_cvs_agg_pipeline_no_exc.py.

Fixed

• The zero division hint for error functions did not properly replace the value with NaN, when the result of the
  operation was -inf instead of inf.

v0.8.0 - Some small quality of life things

[0.8.0] - 2024-09-13

Scientific Changes

• BREAKING: Using "high precision" standard value of gravity everywhere (9.80665 m/s^2).
  Before we were using 9.81 m/s^2 during the loading of MobiliseD matlab files and when calculating the stride length
  validity threshold.
  This should have a small impact on the results of most algorithms, hence, we marked it as a breaking change.

Added

• The apply_aggregations and apply_transformation functions now have an option (on by default) to ignore
  transformations that expect columns that don't exist.
  This should allow the use of the default aggregations and transformations in more situations.

Changed

• The matlab loader now accepts infoForAlgo files that are missing some of the optional fields.

v0.7.0 - [IMPORTANT] Critical bug fix for Impaired Pipeline

[0.7.0] - 2024-09-13

Scientific Changes

• CRITICAL: The default thresholds for both variants of the GSDIonescu algorithms were not correctly adapted to the
  input data beeing in m/s2 instead of g.
  With the old threshold far to many gait sequences were detected.
  This fix should have a substantial positive impact on the performance of the Impaired pipeline.

Fixed

• as_samples now correctly preserves the index of a dataframe.

v0.6.0 - More examples. more flexible loading for ref data

[0.6.0] - 2024-07-26

Added

• Evaluation Examples for CAD and Stride Length (#174)

Changed

• The reference data parser now allows missing information. In case either no turns, no ICs or no strides are available,
  the respective attribute of the reference data is set to None. Before this threw an error.

v0.5.1 - Lower version for pandas bumped

[0.5.1] - 2024-07-24

• Correctly specified minimal pandas version as >=2.2.0

v0.5.0 - **BREAKING-CHANGES** Important turning fixes and finalizing API

[0.5.0] - 2024-07-16

Changed

• IMPORTANT:All algorithms now require the input to be in the body frame.
  This ensures that the correct axis is used in all cases.
  (#170, #171)
• The Coordinate Sytem guide was reworked to cover all different coordinate systems and hopefully prevent common
  mistakes.
  (#170)
• BaseMobilisedPipeline is renamed to GenericMobilisedPipeline. BaseMobilisedPipeline is now the "actual"
  baseclass for Mobilise-D pipelines, without any implementation.
• MobilisedMetaPipeline has been renamed to MobilisedPipelineUniversal.
  This should make the naming more consistent (compare MobilisedPipelineHealthy, MobilisedPipelineImpaired).
• MobilisedPipelineUniversal now has proxy attributes for all the primary results expected in BaseMobilisedPipeline and
  can be used as a drop-in replacement for GenericMobilisedPipeline.
• Gaitmap has been removed as a dependency.
  All used functions are now vendored.
  This should reduce the number of dependency mobgap pulls in.
  (#166)
• Significant performance improvements for GsdAdaptiveIonescu.
  (#167). Thanks @tbonewmy !

Fixed

• IMPORTANT: The TD Elgohary algorithm was using the wrong axis for the calculation of the angle.
  This was fixed and the results should now be correct.
  (#170)
• The turning and stride length algorithm both had an option to be used with prior orientation estimation.
  However, this did not work due to a coordinate system mismatch.
  The Madgwick algorithm has been modified to allow returning data in the correct coordinate system.
  (#170)
• In case no index-col-names are passed to the DatasetFromData class and the data index is just a string and not a
  tuple, a generic column (level_0) is correctly created. (#164)
• The coordinate system was not documented correctly. y and z axis were flipped in the description.

Added

• The Madgwick algorithm is now available in the mobgap package.
  The implementation is copied from gaitmap.
  Note, that its implementation differs slightly from the original Madgwick implementation and it supports the concept
  of body frame data, that only exists in mobgap.
  (#166, #170)
• A bunch new helper tools to perform coordinate transforms.
  (#170)
• Added various array and rotation helpers that were originally imported from gaitmap.
  (#166)
• Added explicit evaluation challenges for the Gait Sequence Detection algorithms.
  (#169)