hdlandmark: Compute individual dynamic prediction of clinical endpoint...

In anthonydevaux/hdlandmark: What the Package Does (One Line, Title Case)

Compute individual dynamic prediction of clinical endpoint using large dimensional longitudinal biomarker history

Description

hdlandmark provides individual survival probabilities using covariates and summaries build on longitudinal data from biomarkers collected over the time. For each biomarker, an ensemble of predictive summaries are computed at the user-specified landmark time tLM. For instance, we use random-effects, level, slope and cumulative level. Then, these summaries and covariates are used as input in several survival prediction methods including: Cox model (his extension with penalty), sparse-Partial Least Square for survival data and random survival forests For each survival prediction method, we provide the individual prediction on horizon time tHor.

Usage

hdlandmark(
  data,
  data.pred = NULL,
  markers,
  tLMs,
  tHors,
  subject,
  time,
  time.event,
  event,
  long.method = c("combine", "GLMM", "MFPC"),
  surv.covar = c("baseline", "LOtLM"),
  cox.submodels = c("autoVar", "allVar"),
  coxnet.submodels = c("opt", "lasso", "ridge"),
  penaFG.submodels = c("GCV", "BIC"),
  spls.submodels = c("opt", "nosparse", "maxsparse"),
  rsf.submodels = c("opt", "noVS", "default"),
  rsf.split = c("logrank", "bs.gradient"),
  cause = 1,
  HW = NULL,
  summaries = c("RE", "score", "pred", "slope", "cumulative"),
  kfolds = 1,
  seed = 1234,
  scaling = FALSE,
  SL.weights = NULL,
  nodesize.grid = NULL,
  mtry.grid = NULL
)

Arguments

data	data.frame object containing longitudinal and survival data
data.pred	(optional) data.frame object for predictions. If missing, data.pred is made using kfolds cross-validation
markers	list containing the modeling of repeated measures for each marker
tLMs	numeric vector of landmark times
tHors	numeric vector of horizon times
subject	variable name in data (and data.pred) that identifies the different subjects
time	variable name in data (and data.pred) which contains time measurements
time.event	variable name in data (and data.pred) which contains time-to-event
event	variable name in data (and data.pred) which contains time-to-event
long.method	character that specifies how to model the longitudinal data. Choices are GLMM for generalized mixed model \insertCitelaird_random-effects_1982hdlandmark, MFPC for multivariate functional principal components \insertCiteyao_functional_2005hdlandmark (works only on continuous markers) or combine for both.
surv.covar	covariates measure at baseline or last observation before landmark time LOtLM
cox.submodels	a character vector containing Cox submodels \insertCitecox_regression_1972hdlandmark. autoVar for Cox with backward variable selection. allVar for Cox with all variables
coxnet.submodels	a character vector containing penalized Cox submodels \insertCitesimon_regularization_2011hdlandmark. opt for tuning the elastic net parameter penalty, lasso for lasso penalty and ridge for ridge penalty.
spls.submodels	a character vector containing Deviance residuals sparse-Partial Least Square sub-methods \insertCitebastien_deviance_2015hdlandmark. opt for tuning sparcity parameter η, nosparse for η = 0 and maxsparse for η = 0.9 \insertCite@see also @chun_sparse_2010hdlandmark
rsf.submodels	a character vector containing random survival forests sub-methods \insertCiteishwaran_random_2008hdlandmark.
rsf.split	a character vector containing the split criterion for random survival forests sub-methods. logrank for log-rank splitting or bs.gradient for gradient-based brier score splitting.
kfolds	number of fold in cross-validation
seed	(optional) seed number
scaling	boolean to scale summaries (default is FALSE)
SL.weights	(optional) allow to compute individual probabilities from a superlearner using numeric vector of weights for each sub-methods

Value

tLMs	landmark time(s)
tHors	horizon time(s)
models	a list for each landmark time(s):

• data.surv input data in survival methods for training (only available on the last fold)

• data.surv.pred input data in survival methods for predicting (only available on the last fold)

• model.surv output object for the selected survival predictive methods (only available on the last fold)

• pred.surv for each horizon time(s), containing the individual probabilities for the selected survival predictive methods

• AUC list of horizon time(s) containing AUC for each fold for the selected survival predictive methods

• BS list of horizon time(s) containing BS for each fold for the selected survival predictive methods

long.method	method(s) used to modeling the biomarkers
surv.methods	method(s) used to compute the individual survival prediction
models.name	name of survival prediction methods
kfolds	number of folds

Author(s)

Anthony Devaux ([e-mail]) (maintener), Robin Genuer and Cécile Proust-Lima

References

\insertAllCited

Examples

## Not run: 

data(pbc2)

# Formula for the modeling of the biomarkers using splines
serBilir = list(model = list(fixed = serBilir ~ year,
                             random = ~ year,
                             subject = "id"),
                deriv = list(fixed = ~ 1,
                             indFixed = 2,
                             random = ~ 1,
                             indRandom = 2))

serChol = list(model = list(fixed = serChol ~ year + I(year^2),
                            random = ~ year + I(year^2),
                            subject = "id"),
               deriv = list(fixed = ~ I(2*year),
                            indFixed = c(2,3),
                            random = ~ I(2*year),
                            indRandom = c(2,3)))

albumin = list(model = list(fixed = albumin ~ year,
                            random = ~ year,
                            subject = "id"),
               deriv = list(fixed = ~ 1,
                            indFixed = 2,
                            random = ~ 1,
                            indRandom = 2))

alkaline = list(model = list(fixed = alkaline ~ year,
                             random = ~ year,
                             subject = "id"),
                deriv = list(fixed = ~ 1,
                             indFixed = 2,
                             random = ~ 1,
                             indRandom = 2))

SGOT = list(model = list(fixed = SGOT ~ year,
                         random = ~ year,
                         subject = "id"),
            deriv = list(fixed = ~ 1,
                         indFixed = 2,
                         random = ~ 1,
                         indRandom = 2))

platelets = list(model = list(fixed = platelets ~ year + I(year^2),
                              random = ~ year + I(year^2),
                              subject = "id"),
                 deriv = list(fixed = ~ I(2*year),
                              indFixed = c(2,3),
                              random = ~ I(2*year),
                              indRandom = c(2,3)))

prothrombin = list(model = list(fixed = prothrombin ~ year,
                                random = ~ year,
                                subject = "id"),
                   deriv = list(fixed = ~ 1,
                                indFixed = 2,
                                random = ~ 1,
                                indRandom = 2))

ascites = list(model = ascites ~ year + (1 + year|id),
               deriv = list(fixed = ~ 1,
                            indFixed = 2,
                            random = ~ 1,
                            indRandom = 2))

hepatomegaly = list(model = hepatomegaly ~ year + (1 + year|id),
                    deriv = list(fixed = ~ 1,
                                 indFixed = 2,
                                 random = ~ 1,
                                 indRandom = 2))

spiders = list(model = spiders ~ year + (1 + year|id),
               deriv = list(fixed = ~ 1,
                            indFixed = 2,
                            random = ~ 1,
                            indRandom = 2))

edema = list(model = edema ~ year + (1 + year|id),
             deriv = list(fixed = ~ 1,
                          indFixed = 2,
                          random = ~ 1,
                          indRandom = 2))

marker <- list(serBilir = serBilir, serChol = serChol, albumin = albumin,
               alkaline = alkaline, SGOT = SGOT, platelets = platelets,
               prothrombin = prothrombin, ascites = ascites,
               hepatomegaly = hepatomegaly, spiders = spiders,
               edema = edema)

# compute hdlandmark methodology
hdlandmark.res <- hdlandmark(data = pbc2, data.pred = pbc2, markers = marker,
                             tLMs = 4, tHors = 3,
                             subject = "id", time = "year", time.event = "years", event = "status2",
                             long.method = "GLMM", surv.covar = "baseline",
                             cox.submodels = "allVar",
                             coxnet.submodels = "lasso",
                             spls.submodels = "nosparse",
                             rsf.submodels = "default",
                             rsf.split = c("logrank"))

# get individual predictions for each method
hdlandmark.res$models[[`4`]]$pred.surv$`3`

## End(Not run)

anthonydevaux/hdlandmark documentation built on Jan. 11, 2023, 8:01 a.m.