Mapping Common Data Model data tables to an HDF5 file for reproducible machine learning workflows

Abstract

Synthetic inpatient claims data in the OMOP Common Data Model were mapped to matrices in a HDF5 file. The content of the HDF5 were further manipulated to build a matrix for a 30-day post discharge readmission model. A random forest model was built to predict a patient’s 30-day readmission risk. While the model’s results were not predictive (AUC = 0.53) the modeling approach and pipeline can be applied to data in the Common Data Model (version 5.0). This opens up the possibility of rigorously comparing predictive performance of readmission models across different datasets.