In novasmedley/gbmSpm: Find predictive sequential patterns for clinical glioblastoma patients

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

The package gbmSPM can be used to do the exact analysis pipeline used in the paper,

Smedley, Nova F., Benjamin M. Ellingson, Timothy F. Cloughesy, and William Hsu. "Longitudinal Patterns in Clinical and Imaging Measurements Predict Residual Survival in Glioblastoma Patients." Scientific reports 8, no. 1 (2018): 14429.

See also Supplemental Materials.

Vignette Info

This vignette shows how temporal features and patient covariates are used in predicing residual survival. This depends on the glmnet and caret R packages, uses dummy patient data, and follows the other vignette: "Generate sequential patterns."

Example

1. Set the cross-validation and logistic regression parameters

   • seed - for data partitioning
   • nFolds - number of cross-validation folds
   • cvReps - number of times to repeat cross-validation
   • metric - for selecting hyperparameters
   • lasso - whether to use lasso regularization (lamba)
   • llength - number of lambas to consider
   • lmax - the max. lamba to consider

   ```r library(gbmSpm)

   seed <- 9 nFolds <- 2 cvReps <- 2 metric <- 'rocAUC' lasso <- TRUE llength <- 100 lmax <- 0.2 ```

2. Set the parameters used to generate features

   Previously, we found temporal features via sequential pattern mining (SPM) with arulesSequences and placed the patterns in "~/gbm_spm_example". We will also put logit results there as well.

   ```r

   previously created spm patterns parameters

   tType <- 'rate' maxgap <- 60 maxlength <- 2 outdir <- '~/gbm_spm_example' dataFolder <- 'sup0.4g60l2z2' # previously created spm patterns dataFile <- file.path(dataFolder, 'featureVectors_rateChange.rds') # data input file

   output dirs for logit models

   prefix <- 'logits' logitFolder <- file.path(outdir, prefix, dataFolder) ifelse(!dir.exists(logitFolder), dir.create(logitFolder, recursive = T), F)

   expected ids of train and test, if doesn't exist, will create it

   dataPartitions <- file.path(outdir, 'train_test_partitions.rds')

   input path, aka spm data folder

   dataPath <- file.path(outdir,'spm',dataFile)

   ```

   Log the experiment:

   ```r logfn <- file(file.path(logitFolder, paste0(format(Sys.Date(), format="%Y.%m.%d"), 'logit',getVolTypeName(tType),'.log')), open='wt')

   sink(logfn, type='output', split=T)

   cat(format(Sys.Date(), format="%Y.%m.%d"), '\n') cat('pattern dataset:', dataFile, '\n') cat('tumor vol variable type: ', getVolTypeName(tType), '\n') cat('seed: ', seed, '\n') cat('nFolds: ', nFolds, '\n') cat('cvReps: ', cvReps, '\n')

   if (lasso) { cat('lambda length: ', llength, '\n\n') cat('lambda max: ', lmax, '\n\n') } ```

3. Create partitions on cleaned data

   Data partitioning depends on the fully cleaned dataset and ensures the exact same clinical visits are maintained in training or testing, regardless of any further feature engineering methods.

   Load original data: ```r data("fake_data") names(fake_data) colnames(fake_data$person) colnames(fake_data$demo) colnames(fake_data$events)

   survData <- fake_data$person fake_demo <- fake_data$demo ```

   Get the pool of visits for partitioning:

   ```r

   'c' tType does not remove initial imaging volumes

   (e.g., rate change needs two visits, the very first one is ignores)

   fake_data$events <- cleanData(fake_data$events, tType = 'c')

   fake_data <- merge(fake_data$events, fake_data$person, by='iois', all.x=T)

   need gender info

   fake_data <- prepDemographics(fake_data, fake_demo)

   get survival labels, these also change

   fake_data <- prepSurvivalLabels(fake_data)

   clinical ids

   fake_data$id <- paste0(fake_data$iois,'.',fake_data$eventID)

   create train and test partitions

   and attempt to maintain gender and survival balance in both parts

   part.ids <- getTrainTestPartition(data = fake_data, database = NULL, personTable = NULL, survData = survData, seed = seed, verbose = T) names(part.ids) ```

4. Read in features for prediction task

   Get the features generated from SPM:

   ```r

   features and labels for each clinical visit

   data <- readRDS(dataPath) names <- colnames(data) data <- data.frame(id=row.names(data),data) colnames(data) <- c('id',names) cat('...overall samples: ', nrow(data), '\n') ```

   Convert tumor laterality and tumor location to dummy variables: r data <- prepLaterality(data) data <- prepLocation(data)

   Remove clinical visits with no history of length max. gap: r data <- removeVisits(data, maxgap = maxgap, maxlength = maxlength, tType = tType, save = F, outDir = logitFolder)

   Remove unwanted features and labels that should not be in training data for glmnet:

   r labels <- getClassLabels() needToRemove <- c('id','iois','eventID', # remove ids labels, # remove labels 'IDH1') # not interested

5. Logistic regression

   Set labels and formula: ```r

   ln <- labels[1] # just do one ln

   form <- as.formula(paste0(ln,'~.')) form ```

   Set training data and do cross-validation with refitting with the best performing lambda:

   ```r data$id <- as.character(data$id) data <- data[data$id %in% part.ids$train,]

   aucs <- runCV(data = data, formula = form, lasso = TRUE, llength = llength, lmax = lmax, labelName = ln, needToRemove = needToRemove, metric = metric, seed = seed, folds = nFolds, cvReps = cvReps, createModelMatrix = FALSE) ```

novasmedley/gbmSpm documentation built on May 17, 2019, 10:39 a.m.