NEWS

survdnn 1.0.2

Dropped the remaining purrr and tidyr dependencies. New internal .map_dfr()/.imap_dfr()/.crossing_grid()/.pivot_longer_simple() helpers (R/utils-tidy.R) replace purrr::map_dfr()/imap_dfr()/ pmap_dfr(), tidyr::crossing(), and tidyr::pivot_longer(). tune_survdnn()'s and gridsearch_survdnn()'s config grids and results carry a hidden list-column (the per-config vector of hidden-layer sizes); since basetable::rbindfill() coerces list-columns to NA instead of preserving them, those two row-binding sites use a new .map_rbind() (base R rbind()) instead of .map_dfr(). purrr and tidyr removed from Imports entirely. Requires basetable (>= 1.4.1).

Dropped the dplyr dependency. summarize_cv_survdnn() and plot.survdnn()'s mean-curve overlay now use basetable::aggregate()/ count(); cv_survdnn()'s column selection and gridsearch_survdnn()/ tune_survdnn()'s row-binding use plain base R. The two group-by aggregations keyed partly on the hidden list-column (per-config vector of hidden-layer sizes) -- tune_survdnn()'s best-config selection and summarize_tune_survdnn() -- use a new internal .aggregate_by_config() (base R split(), since basetable::aggregate() has no list-column grouping support, and dplyr's list-column grouping via vctrs equality needed a structural stand-in: elements collapsed to a string key). rsample::vfold_cv(strata = ...)'s external-vector tidyselect now goes through tidyselect::all_of() directly instead of dplyr::all_of().

Milestone release: the package's methods paper is now published in The R Journal (El Badisy 2026, doi:10.32614/RJ-2026-008), cited in DESCRIPTION's Description field and a matching inst/CITATION file.
Standardized the maintainer's family-name casing to "El Badisy" (title case) in Authors@R, matching the rest of the package suite. Daniel Falbel remains listed as a contributor (ctb).

Hardened \\donttest{} examples that train models by guarding them with torch::torch_is_installed() so checks pass on systems without Torch.
Updated examples to load the veteran dataset explicitly via survival::veteran (instead of data(veteran, ...)).
Added explicit verbose support to cv_survdnn() and tune_survdnn(), with clearer and consistent progress messages across fit/cv/tune workflows.
Added optional .threads argument to survdnn(), cv_survdnn(), and tune_survdnn() to control Torch CPU thread count explicitly when needed.
Regenerated documentation (man/*.Rd) to reflect example updates.
Removed internal-tool references from release notes and submission notes.

Fixed and stabilized loss implementations for AFT and CoxTime models, ensuring correct handling of time scaling, parameter learning, and numerical stability.
Corrected and harmonized prediction methods (predict.survdnn) across all supported losses (Cox, Cox L2, AFT, CoxTime), including survival and risk predictions.
Improved internal consistency checks to ensure valid survival probabilities (monotonicity and bounds).
Updated and expanded unit tests to cover prediction behavior, edge cases, and reproducibility.
Regenerated Rd documentation to fully document all function arguments and remove previous documentation warnings.

Added full support for training control mechanisms, including early stopping callbacks and complete loss tracking across epochs.
Introduced plot_loss() to visualize training loss trajectories and diagnose convergence or instability.
Centralized reproducibility control via the .seed argument in survdnn(), synchronizing both R and Torch random number generators.
Expanded optimizer support to include Adam, AdamW, SGD, RMSprop, and Adagrad, with customizable optimizer arguments.
Enhanced prediction methods to robustly support linear predictors, survival probabilities, and cumulative risk across all supported loss functions.
Added explicit and user-controllable missing-data handling (na_action = "omit" or "fail"), with informative messages.

Improved handling of formulas using Surv(...) ~ . in prediction and evaluation.
Improved printing and summary methods for fitted survdnn objects.
Expanded unit test coverage, including optimizers, plotting utilities, and missing-data edge cases.

Fixed inconsistencies in prediction and evaluation when formulas used . expansion.

Removed automatic torch::install_torch() on load:

The package no longer downloads or installs Torch libraries automatically when loaded. The .onLoad() function now performs only a silent availability check, and .onAttach() displays an informative message instructing users to manually run torch::install_torch() when necessary.

This update ensures full compliance with CRAN policies that forbid modification of user environments or network activity during package load.
Updated startup messages for clearer user guidance.
Internal documentation updates and version bump for CRAN resubmission.

Added conditional test skipping: tests and examples now use skip_if_not(torch_is_installed()) and skip_on_cran() to avoid failures on systems where Torch is not available (thanks to @dfalbel for the PR).
Regenerated documentation (RoxygenNote: 7.3.3) and updated man pages.
Minor internal consistency fixes and CI check updates.

First public release of survdnn.

survdnn(): Fit deep learning survival models using a formula interface.
Supported loss functions:
- Cox partial likelihood ("cox")
- L2-penalized Cox ("cox_l2")
- Time-dependent Cox ("coxtime")
- Accelerated Failure Time ("aft")
Cross-validation via cv_survdnn().
Hyperparameter tuning with tune_survdnn().
Survival probability prediction and curve plotting.
Evaluation metrics: Concordance index (C-index), Brier score, and Integrated Brier Score (IBS).

CRAN submission prepared, including README, documentation, and automated tests.