Skip to contents

Changelog

Source: NEWS.md

ospsuite.parameteridentification 2.2.0

Breaking changes

  • Minimum required R version is now 4.4, consistent with ospsuite.
  • residualWeightingMethod values "std" and "mean" have been removed from residualWeightingOptions. Use outputMapping$scaling = "log" for proportional error handling or "error" for inverse-variance weighting.

Major changes

  • PKOutputMapping adds PK-metric optimization to ParameterIdentification: fit one or more parameters to target PK metrics (C_max, AUC_tEnd, etc.) supplied as scalar values or computed from observed DataSet objects. PK-metric mode and observed time-series mode (via PIOutputMapping) are mutually exclusive within a single ParameterIdentification run.
  • PIConfiguration active bindings (objectiveFunctionOptions, algorithmOptions, ciOptions) now validate input at assignment time, warn on unknown keys, and merge partial lists with current settings. Changing algorithm or ciMethod resets the corresponding options and emits a message (#228).
  • ParameterIdentification$plotResults() migrated from the soft-deprecated tlf-based ospsuite plotting functions (plotIndividualTimeProfile(), plotObservedVsSimulated(), plotResidualsVsTime()) to the new ospsuite.plots-based equivalents (plotTimeProfile(), plotPredictedVsObserved(), plotResidualsVsCovariate()). The DefaultPlotConfiguration object is no longer used; axis scales are derived directly from each PIOutputMapping$scaling, and the residual sub-plot now matches the mapping’s scale instead of being hard-coded to linear. Visual output of plotResults() changes accordingly.
  • Sub-plot composition in plotResults() switched from ospsuite::plotGrid() to patchwork::wrap_plots(); the returned objects are now patchwork objects rather than the previous ospsuite plot-grid objects.

Minor improvements and bug fixes

  • CIOptions_hessian gains r and d options to tune the post-hoc Hessian CI step. r controls the number of iterations (minimum 2, default 4) and d the fractional step size (default 0.1). Reducing r lowers the number of objective function evaluations at the cost of accuracy (#215).
  • residualWeightingMethod = "error" now computes arithmetic SD from geometric standard deviation without approximation. A warning is issued when any error values are invalid and fall back to unit weights (#255).
  • Re-enabled plotOFVProfiles() for visualizing OFV profiles produced by ParameterIdentification$calculateOFVProfiles() (#91).
  • New Imports: patchwork (used to compose the sub-plots produced by plotResults()).
  • ParameterIdentification$estimateCI() now resets the objective function evaluation counter before each bootstrap iteration and each profile likelihood step, so printEvaluationFeedback output restarts from 1 for every sub-optimization (#238).
  • ParameterIdentification now converts observed and simulated data to OSPSuite base units before computing residuals, ensuring consistent and reproducible OFV values (#229, #237).
  • PIResult$toDataFrame() now returns one row per parameter path for grouped PIParameters, instead of only the first path (#230).
  • Removed clearOutputIntervals() call from ParameterIdentification initialization, which could lead to wrong simulation results when events are triggered in time intervals without observed data (#226).

ospsuite.parameteridentification 2.1.1

Breaking Changes

  • Default CI options lists were renamed to CIOptions_hessian and CIOptions_bootstrap for consistency (#220).

Minor improvements and bug fixes

  • Automatically report one-sided confidence intervals when parameter estimate falls outside the bootstrap CI in skewed distributions. In the result, the opposite bound is set to NA and ciType is set to one-sided (#217).
  • Allow dataSet with a single observation (#221)

ospsuite.parameteridentification 2.1.0

Breaking changes

  • ´ospsuite.parameteridentificationnow requiresospsuite.utils` version >= 1.7.0.
  • ´ospsuite.parameteridentificationnow requiresospsuite` version >= 12.2.0.
  • Optimization and CI results are now returned as a PIResult object instead of a list. This provides a unified structure and new helper methods for summaries and export (#196).

Major changes

  • Optimization backend refactored for improved robustness. No changes to the user interface (#161, #186).
  • PIOutputMapping: new $setDataWeights() method for assigning weights to observed data sets and individual data points (#178).
  • New confidence interval methods PL (profile likelihood) and bootstrap are now supported via PIConfiguration (#167, #186, #188, #189, #190).
  • New PIResult supports $print(), $toDataFrame(), and $toList() methods for summaries, export, and diagnostics (#196).

Minor improvements and bug fixes

  • Improved print outputs for all classes (#171).
  • All classes do not inherit from ospsuite.utils::Printable any more (#171).
  • Default settings for CI methods are provided through CIOptions_Hessian, CIOptions_PL, and CIOptions_Bootstrap (#167).
  • Supports ArithmeticStdDev and GeometricStdDev error types passed in the observed DataSet as yErrorType. This has an effect when residualWeightingMethod = "error" is set via PIConfiguration$objectiveFunctionOptions (#181).
  • Harmonized CI output naming: sd, se, cv and rse are now used consistently and correctly (#191).
  • Improved Hessian-based confidence interval estimation: covariance matrix scaled for SSR objective functions (#192).
  • Optimization and CI estimation validated against PK-Sim results for the Aciclovir model, confirming correctness of estimates (#193).
  • Confidence interval estimation can be disabled by setting piConfiguration$autoEstimateCI <- FALSE; it can then be run explicitly with ParameterIdentification$estimateCI() (#196).
  • New vignette on confidence intervals, covering available methods, configuration, and result inspection (#198).
  • New vignette on data mapping, including adding data weights and set data transformation (#199).
  • PIResult stores costDetails from the best (minimum modelCost) evaluation rather than the last, ensuring correct output when runs are limited or stop early (#206).

ospsuite.parameteridentification 2.0.2

Major changes

  • ParameterIdentification$gridSearch() andParameterIdentification$calculateOFVProfiles() are made available and refactored for robustness, clarity, and efficiency (#151).

Minor improvements and fixes

  • ParameterIdentification will validate observed data availability in PIOutputMapping during initialization (#145).
  • Cache Simulation ID in PIOutputMapping(#146).
  • PIOutputMapping will attempt to retrieve the molecular weight for unit conversion when adding observed data (#147).
  • ParameterIdentification now differentiates between simulation failures during the first iteration (stopping optimization) and subsequent iterations (returning infinite cost structure) (#148).
  • Simulation failure in gridSearchand calculateOFVProfile won’t break evaluation and return Inf for specific parameters (#153).
  • Robust Hessian epsilon calculation in ParameterIdentification (#160).

ospsuite.parameteridentification 2.0.1

Breaking changes

  • Function getSteadyState has been removed in favor of getSteadyState from the {ospsuite} package (#128).
  • Function validateIsOption has been removed in favor of ospsuite.utils::validateIsOption (#130).

Minor improvements and fixes

  • Fix pkgdown build (#131)
  • Fix bug in simulation ID verification (#138)

ospsuite.parameteridentification 2.0.0

Breaking changes

  • Requires {ospsuite} version 12.0 or later (#98, #114).

  • PIConfiguration now configures the objective function via objectiveFunctionOptions. Users can specify options directly, including objectiveFunctionType, residualWeightingMethod, robustMethod, scaleVar, and linScaleCV (#100).

  • ParameterIdentification$gridSearch() andParameterIdentification$calculateOFVProfiles() functions have been disabled until fixed (#91, #92).

Major changes

  • New calculateCostMetrics() function in ParameterIdentification enhances model evaluation with integrated configuration via PIConfiguration. Settings and defaults are specified in ObjectiveFunctionOptions (#64, #65, #100).

    • objectiveFunctionType: lsq for least squares or m3 for censored data maximum likelihood estimation.
    • residualWeightingMethod: options include none, std normalizes by standard deviation for variable data, mean scales by mean for diverse magnitudes, and error weights by inverse variance for known error data.
    • robustMethod: none for uniform treatment, huber or bisquare for outlier minimization.
    • scaleVar: A boolean indicating whether to scale residuals by the number of observations.
    • linScaleCV: Numeric coefficient used to calculate standard deviation for linear scaling, applied to lloq values when m3method is used.
    • logScaleSD: Numeric standard deviation for logarithmic scaling, applied to lloq values when m3method is used.

  • New error-calculation vignette, explaining error model methodologies within the ospsuite.parameteridentification package. The document elaborates on the lsq (Least Squares Error) and m3 (Extended Least Squares Error for censored data) error models, along with advanced customization options for error modeling. This resource aids users in refining their parameter identification processes. (#102, #111).

  • New optimization-algorithms vignette, introducing algorithms available for parameter estimation and offering insights on their optimal application scenarios (#104, #111).

  • New user-guide vignette on parameter identification (PI). This vignette provides a comprehensive overview for setting up PI tasks, including defining simulations, specifying parameters to be identified, mapping model outputs to observed data, and configuring optimization tasks. Examples across three complexity levels of models demonstrate the package’s functionality in detail (#48).

  • New plot.modelCost() function for visualizing raw and weighted residuals from modelCost objects (#100).

  • Comprehensive overhaul of documentation, for clarity, comprehensiveness, and ease of navigation for all users (#111).

  • Unit test overhaul and coverage enhancement, increasing from 67% to 85% (#99, #100, #113).

Minor improvements and fixes

  • Enhanced error handling now ensures ParameterIdentification tasks validate the existence of simulation objects referenced by PIParameters and OutputMapping to prevent runtime errors due to missing dependencies (#117, #120).

  • Enhanced calculateCostMetrics() output in ParameterIdentification offers a more detailed results summary for model evaluation. The summary now includes modelCost, minLogProbability, and a costVariables dataframe with scaleFactor, nObservations, M3Contribution, SSR (sum of squared residuals), weightedSSR, normalizedSSR, and robustSSR (#100).

  • README file conversion to .rmd format and subsequent updates (#86, #93).

  • Improved error handling in ParameterIdentification for cases of simulation failure, ensuring consistent and informative error cost structure in output (#66, #70, #100).

  • validateIsOption() ensures user-specified options adhere to defined constraints, enhancing the robustness of user inputs (#100).

  • ParameterIdentification now directly accesses default optimization algorithm options for BOBYQA, HJKB, and DEoptim from their respective packages (nloptr, {dfoptim}, and DEoptim) if not explicitly defined in PIConfiguration (#48, #81).

  • Continuous Integration/Continuous Deployment pipeline improvements (#95, #106, #110)

  • Several bug fixes (#83, #109, #110, #115, #119, #122)

ospsuite.parameteridentification 1.3

Breaking changes

  • The parameter in the PIConfiguration class that is controlling the feedback at each function evaluation is now called printEvaluationFeedback instead of printItera tionFeedback.

Major changes

  • Added new optimization algorithms: the default local algorithm is now an implementation of the BOBYQA algorithm (bounded optimization by quadratic approximation) from the nloptr package; additional local algorithm is HJKB, a bounded implementation of the Hooke-Jeeves derivative-free algorithm from the {dfoptim} package; a global algorithm is DEoptim for differential evolution optimization.

  • FME::modCost() is re-implemented as part of the parameter identification package and used for calculation of residuals.

Minor bug fixes and improvements

  • Calculation of residuals does not fail if observed data contains only one time point.
  • Calculation of the hessian close to the bounds of parameter values is improved.

ospsuite.parameteridentification 1.2

Breaking changes

Minor bug fixes and improvements