A Method for Machine Learning Generation of Realistic Synthetic Datasets for Validating Healthcare Applications

Abstract

ABSTRACTBackgroundDigital health applications can improve quality and effectiveness of healthcare, by offering a number of tools to patients, professionals, and the healthcare system. Introduction of new technologies is not without risk, and digital health applications are often considered a medical device. Assuring their safe operation requires, amongst others, clinical validation, which needs large datasets to test their application in realistic clinical scenarios. Access to such datasets is challenging, due to concerns about patient privacy. Development of synthetic datasets, which will be sufficiently realistic to test digital applications, is seen as a potential alternative, enabling their deployment.ObjectiveThe aim of work was to develop a method for the generation of realistic synthetic datasets, statistically equivalent to real clinical datasets, and demonstrate that Generative Adversarial Network based approach is fit for purpose.MethodA generative adversarial network was implemented and trained, in a series of six experiments, using numerical and categorical variables from three clinically relevant datasets, including ICD-9 and laboratory codes from the MIMIC III dataset. A number of contextual steps provided the success criteria for the synthetic dataset.ResultsThe approach created a synthetic dataset that exhibits very similar statistical characteristics with the real dataset. Pairwise association of variables is very similar. A high degree of Jaccard similarity and a successful K-S test further support this.ConclusionsThe proof of concept of generating realistic synthetic datasets was successful, with the approach showing promise for further work.