Twin

For R 4.0.0, dplyr 1.0.0, and tibble 3.0.0.

Baldy

Fixes for glm and R 3.6.
Coincides with the CRAN 2.4.0 release.

Superior

• healthcareai now depends on recipes 0.1.4 and caret 6.0.81. You will need these versions or later. Various hidden changes were made to be compatible with these packages' latest breaking changes.
• bagimpute in prep_data now accepts bag_trees to specify the number of trees. This is updated to be compatible with recipes 0.1.4.
• Local loaded healthcareai library versions now are saved to model objects.

Lone Peak

Added

• Support for models of outcomes with more than two classes (multiclass).
• Explore a model's logic with explore. Make counterfactual predictions across the most-important features in a model to see how those features influence predicted outcomes.
  • plot method to visualize a model's logic.
• Identify opportunities to improve a patient's outcome with Patient Impact Predictor, pip. Carefully specify variables and alternative values that exert causal influence on outcomes; then get recommended actions for a given patient with expected outcomes given the actions.
• Predict outcome groups based on how bad false alarms are relative to missed detections (outcome_groups argument to predict).
• Group predictions into risk groups using the risk_groups argument to predict.
  • plot support for outcome- and risk-group predictions.
• Get thresholds to split outcome classes to optimize various performance metrics with get_thresholds.
  • plot method to compare performance across metrics at various thresholds.
• split_train_test can keep multiple observations of an individual in the same split via the grouping_col argument.
• Replace values that represent missingness but have been interpreted by R as strings with NA with make_na.
  • If missingness finds any such strings it issues a warning with code that can be used to do the replacement.
• Add counts to factor levels with rename_with_counts.
• summary.missingness method for wide datasets with missingness in many columns.

Changed

• In prep_data, trigonometric transformations make circular features out of dates and times for more informative features in less-wide data frames.
• Fixed AUPR in plot.model_class and summary.model_class.
• Can specify performance metric to optimize in machine_learn.
• missingness is faster.
• Regression prediction plots are plotted at 1:1 aspect ratio.
• add_best_levels works in deployment even if none of the columns to be created are present in the deployment observations.
• prep_data can handle logical features.
• outcome doesn't need to be re-declared in model training if it was specified in data prep.

Removed

• No longer support training models on un-prepped data.
• No longer support wrapping caret-trained models into a model_list.

White Baldy

Added

• pima_meds dataset to accompany pima_diabetes

Changed

• Fixed a bug where predicting on XGBoost could fail if column order was different from training

Snowmass Mountain

Added

• Identify values of high-cardinality variables that will make good features, even with multiple values per observation with add_best_levels and get_best_levels.
• glmnet for regularized linear and logistic regression.
• interpret and plot.interpret to extract glmnet estimates.
• XGBoost for regression and classification models.
• variable_importance returns random forest or xgboost importances, whichever model performs better.

Changed

• predict can now write an extensive log file, and if that option is activated, as in production, predict is a safe function that always completes; if there is an error, it returns a zero-row data frame that is otherwise the same as what would have been returned (provided prep_data or machine_learn was used).
• Control how low variance must be to remove columns by providing a numeric value to the remove_near_zero_variance argument of prep_data.
• Fixed bug in missingness that caused very small values to round to zero.
• Messages about time required for model training are improved.
• separate_drgs returns NA for complication when the DRG is missing.
• Removed some redundent training data from model_list objects.
• methods is attached on attaching the package so that scripts operate the same in Rscript, R GUI, and R Studio.
• Minor changes to maintain compatibility with ggplot2, broom, and recipes.

Removed

• Removed support for k-nearest neighbors
• Remove support for maxstat splitting rule in random forests

Kings Peak

This release introduces a new package architecture and API. Please visit https://docs.healthcare.ai to learn more.

Version 2.0 BETA

This is a beta version of a new package architecture. Please provide feedback here.

Final Fir

This is a patch to conform to CRAN policy about not writing to any user directory other than the working directory; the package had previously had tests that wrote to a sqlite file in inst/extdata.

v2.0 is coming soon, which introduces an entirely new architecture and breaking changes across the package. A startup message warning about these coming changes has been added.

Prolific Pines

• Maintains compatibility with new versions of ranger and caret.
• Improves hyperparameter tuning in RandomForestDevelopment