OMOP - Patient Level Prediction

This tutorial will illustrate how to use the PatientLevelPrediction package created by the OHDSI community and described here: https://ohdsi.github.io/PatientLevelPrediction/. Researchers can use this package to predict risk of specific outcomes for persons in a target cohort. The prediction results can be analyzed programmatically or in an R Shiny application.

The Pharmetrics dataset (containing real patient claims data) will be used in this tutorial to ensure that correlation between risk of outcome and presence of outcome is consistent (Synthetic data may not produce consistent relationships between covariates). Any person-level results should not be shared with individuals who are not active OHDSI Lab users.

We’ll need to install one new package: PatientLevelPrediction

renv::install("OHDSI/PatientLevelPrediction")

We’ll generate a cohort from ATLAS as we did in vignette 3, except now we need two cohorts: a target cohort (i.e. the population we’re analyzing), 4675 (Persons with type 2 diabetes and prescribed metformin within 0-30 days after their type 2 diabetes diagnosis) and an outcome cohort (i.e. the outcome we’re predicting risk for) 4681 (Persons with congestive heart failure).

# Packages =====================================================================
library(PatientLevelPrediction)
library(DatabaseConnector)
library(DatabaseConnector)
library(keyring)
library(CohortGenerator)
library(tidyverse)

# DB and ATLAS Connections =====================================================
atlas_url = "https://atlas.roux-ohdsi-prod.aws.northeastern.edu/WebAPI"
cdm_schema = "omop_cdm_53_pmtx_202203"
write_schema = paste0("work_", keyring::key_get("db_username"))

# Setting connectionDetails for later use
Sys.setenv("DATABASECONNECTOR_JAR_FOLDER" = "Insert path to JDBC driver")

connectionDetails <- createConnectionDetails(
    dbms = "redshift",
    server = "ohdsi-lab-redshift-cluster-prod.clsyktjhufn7.us-east-1.redshift.amazonaws.com/ohdsi_lab",
    port = 5439,
    user = keyring::key_get("db_username"),
    password = keyring::key_get("db_password"))

# Connecting to the database for the purpose of setting the following options
con =  DatabaseConnector::connect(connectionDetails)

options(con.default.value = con)
options(schema.default.value = cdm_schema)
options(write_schema.default.value = write_schema)

# Connecting to ATLAS
ROhdsiWebApi::authorizeWebApi(
    atlas_url,
    authMethod = "db",
    webApiUsername = keyring::key_get("atlas_username"),
    webApiPassword = keyring::key_get("atlas_password"))

# Choosing an ATLAS cohort definition
cohortIds <- c(4675, 4681)

# Pulling the cohort definition from ATLAS
cohortDefinitionSet <- ROhdsiWebApi::exportCohortDefinitionSet(
    baseUrl = atlas_url,
    cohortIds = cohortIds)

# Setting naming convention for cohort tables
cohortTableNames <- getCohortTableNames(cohortTable = "cohort")

# Creating cohort tables based on naming convention
createCohortTables(
    connectionDetails = connectionDetails,
    cohortTableNames = cohortTableNames,
    cohortDatabaseSchema = write_schema)

# Generating the cohort population from the Pharmetrics dataset
cohortsGenerated <- generateCohortSet(
    connectionDetails = connectionDetails,
    cdmDatabaseSchema = cdm_schema,
    cohortDatabaseSchema = write_schema,
    cohortTableNames = cohortTableNames,
    cohortDefinitionSet = cohortDefinitionSet)

Now, we need to create the settings for the prediction model. First we set up the database and cohort details.

databaseDetails <- PatientLevelPrediction::createDatabaseDetails(
  connectionDetails = connectionDetails,
  cdmDatabaseSchema = cdm_schema,
  cdmDatabaseName = cdm_schema,
  cdmDatabaseId = "pmtx",
  cohortDatabaseSchema = write_schema,
  cohortTable = "cohort",
  outcomeDatabaseSchema = write_schema,
  outcomeTable = "cohort",
  targetId = 4675,
  outcomeIds = 4681)

Next, we create covariates using Feature Extraction as shown in vignette 6. This may take some time to run.

plpData <- PatientLevelPrediction::getPlpData(
  databaseDetails = databaseDetails,
  covariateSettings = FeatureExtraction::createDefaultCovariateSettings())

So we don’t have to run the above again, we save plpData as an object and then load it again.

#save plpData
PatientLevelPrediction::savePlpData(plpData, "plpData")

#load plpData
plpData <- PatientLevelPrediction::loadPlpData("plpData")

Next, we set our population settings, choosing which persons in our cohort to include/remove from the prediction.

populationSettings <- PatientLevelPrediction::createStudyPopulationSettings(
  binary = TRUE,
  firstExposureOnly = FALSE,
  washoutPeriod = 0,
  removeSubjectsWithPriorOutcome = FALSE,
  priorOutcomeLookback = 99999,
  requireTimeAtRisk = TRUE,
  minTimeAtRisk = 0,
  riskWindowStart = 0,
  startAnchor = 'cohort start',
  riskWindowEnd = 365,
  endAnchor = 'cohort start')

Here,we select the type of prediction model we want to use. In this example, we will use lasso logistic regression.

lr_model <- PatientLevelPrediction::setLassoLogisticRegression()

Now, we’re ready to run the model (and set a few more settings). Depending on the size of our cohort and other settings, this may take a while (when I ran it, it took 2 hours). Feel free to create more restrictive cohorts or edit any of the prediction settings.

lr_results <- PatientLevelPrediction::runPlp( 
  plpData = plpdata, 
  outcomeId = 4681,
  analysisId = 'demo', 
  analysisName = 'run plp demo', 
  populationSettings = populationSettings, 
  splitSettings = PatientLevelPrediction::createDefaultSplitSetting(
    type = "time",
    testFraction = 0.25,
    nfold = 2), 
  sampleSettings = PatientLevelPrediction::createSampleSettings(),
  preprocessSettings = PatientLevelPrediction::createPreprocessSettings(
    minFraction = 0, 
    normalize = T), 
  modelSettings = lr_model, 
  executeSettings = PatientLevelPrediction::createDefaultExecuteSettings(), 
  saveDirectory = "insert path to directory you want to save your results to")

Again, because the above code takes so long to run, let’s save the results as an object for quick loading.

#save results
PatientLevelPrediction::savePlpResult(lr_results, "demo_results")

#load results
lr_results <- PatientLevelPrediction::loadPlpResult("demo_results")

Now that we have run the prediction model, let’s view the results in two of ways. First, to launch the R Shiny application, we run the following code.

PatientLevelPrediction::viewPlp(lr_results)

Finally, let’s view the results in tabular form. Once you run the following code, note that the value column contains the risk that that person had of experiencing the outcome. If we calculated the average of the value column for each outcome (0 meaning no outcome, and 1 meaning with outcome), we could judge whether the model we used was successful in prediction.

results <- PatientLevelPrediction::loadPlpResult("insert path to plpResult directory It should have saved within the directory you set above.")
View(results$prediction)