Is there a way to use the exact measurement values in PLP?

[arani11]

We used the PLP package to develop a model and weight came out as the top predictor. However, our understanding is that in the PLP this is used as a binary variable (yes/ no) for whether weight was recorded or not. We are interested in using weight as a continuous variable for our prediction model using the PLP package and wondered if there was a way to do this?

[egillax]

Hi @arani11 ,

This depends on how you specify your covariateSettings. For example to extract age, sex and measurement values I'd use:

plpData <- getPlpData(
  databaseDetails = myDatabaseDetails,
  covariateSettings = FeatureExtraction::createCovariateSettings(
    useDemographicsAge = TRUE,
    useDemographicsGender = TRUE,
    useMeasurementValueLongTerm = TRUE
  ),
  restrictPlpDataSettings = createRestrictPlpDataSettings())

Or alternatively if you're using runMultiplePlp the covariateSettings are defined the same way in the modelDesign

modelDesign <- createModelDesign(
  targetId = 1,
  outcomeId = 2,
  restrictPlpDataSettings = createRestrictPlpDataSettings(),
  populationSettings = popSettings,
  covariateSettings = FeatureExtraction::createCovariateSettings(
    useDemographicsAge = TRUE,
    useDemographicsGender = TRUE,
    useMeasurementValueLongTerm = TRUE
  ),
  preprocessSettings = createPreprocessSettings(),
  splitSettings = createDefaultSplitSetting(splitSeed = 42),
  modelSettings = setLassoLogisticRegression(seed=42)
)

Here I use useMeasurementValueLongTerm to use the measurement values themselves (in last year) instead of only the binary indicators (which use useMeasurementLongTerm). If you're using default covariateSettings I think it uses the binary indicator ones.

Be careful though, missing measurement values will be treated as zero. You would need to use some kind of imputation to handle that which is possible using the featureEngineering functionality. Recently createStratifiedImputationSettings function was added as a feature engineering method, which is the first imputation method built into the package. If you need any help with that let me know and I can throw together a short example. Either with that one or a simple mean imputation.

[egillax]

There will be more imputation methods added to 6.4. Tracked with #461. Closing this issue in favour of that one.

[arani11]

Thanks!

…

________________________________ From: Egill Axfjord Fridgeirsson ***@***.***> Sent: Monday, December 9, 2024 9:14 AM To: OHDSI/PatientLevelPrediction ***@***.***> Cc: Arani Vivekanantham ***@***.***>; Mention ***@***.***> Subject: Re: [OHDSI/PatientLevelPrediction] Is there a way to use the exact measurement values in PLP? (Issue #459) There will be more imputation methods added to 6.4. Tracked with #461<#461>. Closing this issue in favour of that one. — Reply to this email directly, view it on GitHub<#459 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AZ5I3PPS4VDFVPUCBXPPPAT2EVNQXAVCNFSM6AAAAABI67XH4OVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKMRXGM2TAMBSG4>. You are receiving this because you were mentioned.Message ID: ***@***.***>