- Analytics use case(s): Patient-Level Prediction
- Study type: Methods Research
- Tags: -
- Study lead: Egill Fridgeirsson
- Study lead forums tag: -
- Study start date: -
- Study end date: -
- Protocol: -
- Publications: -
- Results explorer: -
An OHDSI study package that adds signal-adaptive ridge (SAR) and signal-adaptive LASSO (SAL) logistic regression to PatientLevelPrediction and compares them with ridge and LASSO on a set of prediction tasks across OMOP CDM databases. In SAR, each covariate's ridge penalty is scaled by a weight computed from the training rows of each fit: its prevalence, its information and its marginal association with the outcome under the intercept-only model. The weight shape and the penalty are chosen by cross-validation, with the weights recomputed inside every fold. SAL uses the same weights with a LASSO penalty.
- Running the study at a site: see README_PRODUCTION.md and
extras/CodeToRun.R. - Cohorts: Circe definitions in
inst/cohorts(the rendered SQL is shipped, so no Java or CirceR is needed). - What leaves a site: aggregate metrics and counts only, with small counts censored (
exportResults()).
Status: in development; not yet for production use.