Skip to contents

Changelog

Source: NEWS.md

martini 0.7.0

breaking changes

  • default for id column changed to USUBJID according to ADaM implementation guide

  • build(join = ) default changed from dplyr::inner_join() to dplyr::full_join(), which is equivalent to “adsl” (if included) according to CDISC standards

  • default for correlation method for prepare_ml()s prep_step_corr is now spearman instead of pearson in consistency with our major use case with random forests. Can be controlled via prepare_ml(corr_method).

  • the prepare_ml() output object structure has changed

    • similar to the data that is provided in its raw form as well as prepped and ready to use for ML, the recipe is now provided in raw and prepped version as well to allow for proper tuning workflows
    • removedcolscolscorr_keep
    • recipe log step now called log_skewness. Info on log transformed values to be extracted from new output slot

  • the default recipe itself has been modified, making use of new custom steps that are adaptations of existing recipes functions

    • step_corr_keep() offers the functionality of recipes::step_corr() with the option to nominate preferred representatives of correlated groups.
    • step_log_skewness() to transform variables exceeding a threshold for skewness before imputation by means of step_impute_knn() for a more robust imputation. Optionally, variables can be transformed back to their original scale right after imputation using step_log_skewness_undo().
    • step_other2() is pretty much identical with recipes::step_other() except its behavior in case of a single class that does not meet the required minimum size as defined by threshold: while the original version renames the class and silently adds novel levels to this class, our version would leave the single class unmodified and raise an error in case a new data set comes with novel levels for that variable. This change was handled manually before and does not change results.

  • argument prep_recipe is deprecated in favor of custom_recipe for consistency

  • example data sets in the package are updated according to new output structure described above

  • vars_count deprecated in favor of vars_no_trafo: count variables are no longer guessed and automatically excluded from transformation, but need to be specified by the user.

  • vars_ordinalscore deprecated.

major changes

  • the output of prepare_ml() now comes with its own class martini_ml and corresponding print method

  • the feature matrix check functionality has been extended. [check_freq()] is now part of a new function [check_feature()] that can e.g. also check for outliers in numeric variables. It is run by default in [prepare_ml()], but recommended to run prior to calling [prepare_ml()].

minor changes

  • [adam_spec()]’s option to include data sets in ADaM format that are not part of the limited internal library has been extended to data sets following occds data structure: adam_spec(add_occds = ..., add_bds = ...).

documentation

martini 0.6.4

major changes

  • to accommodate for handling of adsl data sets in ADaM 2.0 format, the structure of fct_levels entry was changed to name value pairs

  • added function adjust_spec_filter(), removing append argument from adjust_spec()

  • adjust_spec() now provides extensive checks of user defined modifications to a spec object. Part of them relies on the data being available, which is why we strongly recommend to attach data in adam_spec(), if possible, which is why the default for attach_data in adam_spec() is now TRUE

martini 0.6.3

  • preparations for bay-open: added code of conduct, updated license and added code owner and contributing information

martini 0.6.2

  • removed ´corrr´ package dependency
  • introducing github actions

martini 0.6.1

  • updated output of correlation handling (removals$cols structure and addition of alternative labels in dictionary)

martini 0.6.0

  • in rare cases, samples could end up being removed from training data set due to incomplete imputation from recipes::step_impute_knn(). For this edge case, additional imputation steps were added to ensure full data set usage for training (recipes::step_impute_median() for numerics and recipes::step_impute_mode() for factors.)

  • update example objects, since prepare_ml() output object

    • has new entry high_corr for more transparency on feature dropping for correlation (recipes::step_corr())
    • no longer contains redundant slot for split object

  • variables that are defined in prepare_ml()’s vars_keep_corr but not present in data set are ignored (previously an error was thrown)

  • improved handling of independent data sets for prediction

  • improved test suite

martini 0.5.1

  • Added experimental parameter rm in build() to allow for the preparation of a wide data set suitable for repeated measurement outcomes (one row is a subject at a specific time point)

martini 0.5.0

  • fixed package data sets. Update of {recipes} package introduced NAs in prepared ML data, which were now removed.
  • some refactoring
  • speed improvements in prepare_ml() (affects imputation seed)
  • added some tests

martini 0.4.3

  • update of example data sets
  • minor bug fixes related to incomplete bds data sets (e.g. missing units)

martini 0.4.2

  • improved messaging (wip)
  • updated docu
  • switched to testthat 3e and updated tests
  • some refactoring

martini 0.4.1

Breaking changes

  • change one_hot default to FALSE.

martini 0.4.0

  • `build_bds()´ converts values character values (e.g. AVALC) to either numerics or factors, based on observed values per parameter (e.g. PARAMCD)

martini 0.3.4

  • Added a NEWS.md file to track changes to the package.
  • Dictionary (if available) is updated if variables are added/dropped with adjust_adsl().
  • If data is attached to the spec object, usage of adjust_adsl() and adjust_spec() will update the data_info and filter check attributes shown by the print method