Survival Priority Lasso Learner

Patient outcome prediction based on multi-omics data taking practitioners’ preferences into account. Calls prioritylasso::prioritylasso() from prioritylasso.

Prediction types

This learner returns three prediction types:

1. lp: a vector containing the linear predictors (relative risk scores), where each score corresponds to a specific test observation. Calculated using prioritylasso::predict.prioritylasso().

2. crank: same as lp.

3. distr: a survival matrix in two dimensions, where observations are represented in rows and time points in columns. Calculated using mlr3proba::breslow() where the Breslow estimator is used for computing the baseline hazard.

Initial parameter values

• family is set to "cox" for the Cox survival objective and cannot be changed

• type.measure is set to "deviance" (cross-validation measure) and cannot be changed

• cox.ties is initialized to "breslow" to keep the tie-handling behavior of earlier glmnet versions, and to silence the glmnet v5.0 warning about the upcoming default change to "efron"

Dictionary

This Learner can be instantiated via lrn():

lrn("surv.priority_lasso")

Meta Information

• Task type: “surv”

• Predict Types: “crank”, “distr”, “lp”

• Feature Types: “logical”, “integer”, “numeric”

• Required Packages: mlr3, mlr3proba, prioritylasso

Parameters

Id	Type	Default	Levels	Range
blocks	untyped	-		-
max.coef	untyped	NULL		-
block1.penalization	logical	TRUE	TRUE, FALSE	-
lambda.type	character	lambda.min	lambda.min, lambda.1se	-
standardize	logical	TRUE	TRUE, FALSE	-
nfolds	integer	10		\([3, \infty)\)
foldid	untyped	NULL		-
cvoffset	logical	FALSE	TRUE, FALSE	-
cvoffsetnfolds	integer	10		\([1, \infty)\)
return.x	logical	TRUE	TRUE, FALSE	-
lambda	untyped	NULL		-
grouped	logical	TRUE	TRUE, FALSE	-
trace.it	integer	0		\([0, 1]\)
cox.ties	character	breslow	breslow, efron	-
include.allintercepts	logical	FALSE	TRUE, FALSE	-
use.blocks	untyped	"all"		-
adaptive.order	logical	FALSE	TRUE, FALSE	-

Scope and supported arguments

This learner intentionally exposes a focused subset of training and prediction arguments. It is designed to work well out of the box, without requiring extensive parameter tuning. Some arguments from cv.glmnet(), glmnet(), and predict.prioritylasso() are not included, because they are not consistently supported or forwarded through the full train/predict path (e.g. handling missing test data, or performing relaxed lasso fits).

Please open an issue if there is a need for supporting more learner parameters.

Custom mlr3 parameters

• adaptive.order: if TRUE, the priority order of blocks is estimated from the data following Herrmann et al. (2021), instead of using the user-supplied block order. For each block, a Ridge regression (alpha = 0) is fit using cv.glmnet() on that block alone. The importance of a block is measured by the mean absolute coefficient (MAC) score at the lambda.min value (the lambda giving minimum cross-validation error). A penalty factor of 1 / MAC is then assigned to each block. Blocks are sorted by increasing penalty factor, i.e., blocks with larger MAC (stronger average signal) receive higher priority (come first). Also, the block‑wise penalty factors are attached to the fitted model object as learner$model$block.penalty.factors.

This method is useful when no domain knowledge is available to specify block priority. In this step, data are standardized by default (standardize = TRUE), but this can be overridden by the learner's standardize parameter. lambda.min is always used to derive the block priority. Additional arguments such as nfolds, type.measure, weights and cox.ties (if provided) are forwarded to each block‑wise cv.glmnet() fit. The max.coef parameter, if supplied, it is re‑ordered accordingly to align with the new block order.

This parameter is ignored when fewer than two blocks are provided. It defaults to FALSE for backward compatibility.

References

Simon K, Vindi J, Roman H, Tobias H, Anne-Laure B (2018). “Priority-Lasso: a simple hierarchical approach to the prediction of clinical outcome using multi-omics data.” BMC Bioinformatics, 19. doi:10.1186/s12859-018-2344-6 .

Herrmann, M., Probst, P., Hornung, R., Jurinovic, V., Boulesteix, L. A (2021). “Large-scale benchmark study of survival prediction methods using multi-omics data.” Briefings in Bioinformatics, 22(3), 1–15. doi:10.1093/BIB/BBAA167 .

See also

• Dictionary of Learners: mlr3::mlr_learners.

• as.data.table(mlr_learners) for a table of available Learners in the running session (depending on the loaded packages).

• Chapter in the mlr3book: https://mlr3book.mlr-org.com/chapters/chapter2/data_and_basic_modeling.html#sec-learners

• mlr3learners for a selection of recommended learners.

• mlr3cluster for unsupervised clustering learners.

• mlr3pipelines to combine learners with pre- and postprocessing steps.

• mlr3tuning for tuning of hyperparameters, mlr3tuningspaces for established default tuning spaces.

Author

HarutyunyanLiana

bblodfon

Super classes

mlr3::Learner -> mlr3proba::LearnerSurv -> LearnerSurvPriorityLasso

Methods

Public methods

• LearnerSurvPriorityLasso$new()

• LearnerSurvPriorityLasso$selected_features()

• LearnerSurvPriorityLasso$clone()

Inherited methods

• mlr3::Learner$base_learner()
• mlr3::Learner$configure()
• mlr3::Learner$encapsulate()
• mlr3::Learner$format()
• mlr3::Learner$help()
• mlr3::Learner$predict()
• mlr3::Learner$predict_newdata()
• mlr3::Learner$print()
• mlr3::Learner$reset()
• mlr3::Learner$train()

---

LearnerSurvPriorityLasso$new()

Creates a new instance of this R6 class.

Usage

LearnerSurvPriorityLasso$new()

---

LearnerSurvPriorityLasso$selected_features()

Selected features, i.e. those where the coefficient is non-zero.

Usage

LearnerSurvPriorityLasso$selected_features()

Returns

character().

---

LearnerSurvPriorityLasso$clone()

The objects of this class are cloneable with this method.

Usage

LearnerSurvPriorityLasso$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.

Examples

# Define a Task
task = tsk("grace")

# Create train and test set
ids = partition(task)

# Check task's features
task$feature_names
#> [1] "age"        "los"        "revasc"     "revascdays" "stchange"  
#> [6] "sysbp"     

# Partition features to 2 blocks
blocks = list(bl1 = 1:3, bl2 = 4:6)

# Define learner
learner = lrn("surv.priority_lasso", blocks = blocks, block1.penalization = FALSE,
              lambda.type = "lambda.1se", standardize = TRUE, nfolds = 5)

# Train the learner on the training ids
learner$train(task, row_ids = ids$train)

# Selected features
learner$selected_features()
#> [1] "age"        "los"        "revasc"     "revascdays"

# Make predictions for the test rows
predictions = learner$predict(task, row_ids = ids$test)
predictions
#> 
#> ── <PredictionSurv> for 330 observations: ──────────────────────────────────────
#>  row_ids  time status     crank        lp     distr
#>        1 180.0  FALSE 0.4873027 0.4873027 <list[1]>
#>        6   5.0  FALSE 1.4541773 1.4541773 <list[1]>
#>        8   2.0  FALSE 1.1603463 1.1603463 <list[1]>
#>      ---   ---    ---       ---       ---       ---
#>      995   0.5   TRUE 3.9088136 3.9088136 <list[1]>
#>      996  69.0   TRUE 3.9357878 3.9357878 <list[1]>
#>     1000  15.0  FALSE 3.7133697 3.7133697 <list[1]>

# Score the predictions
predictions$score()
#> surv.cindex 
#>   0.7351317