Fig. 1: An overview of the synthetic EHR data generation benchmarking framework.
From: A Multifaceted benchmarking of synthetic electronic health record generation models
The framework is composed of three phases: (1) a synthetic EHR data generation process, (2) a multifaceted assessment process, and (3) a use case-specific model recommendation process. In Phase 1, given a synthesis paradigm and the real data, we generate multiple (specifically, three in our experiments) synthetic datasets using each data generation model. In Phase 2, each generated synthetic dataset is assessed and assigned a value in terms of each assessment metric. Afterward, all synthetic datasets will be ranked according to their values in terms of each metric. In Phase 3, for each use case, we assign a weight to each metric and convert multiple (specifically, twelve in our experiments) ranking lists of synthetic datasets into one ranking list of models. Finally, the top ranked model for each use case is recommended. EHR: electronic health record.
