Data Balancing with Synthetic Medical Data Generation

Abstract

There are ethical, bureaucratic and operational difficulties in obtaining and using personal health data in the areas that require the use of sensitive health data such as health care planning, clinical trials and research and development studies. The cost and time consuming of obtaining data from clinical and field studies, especially the restrictions on the security of electronic personal health records and personal data privacy, necessitate the production of synthetic data as close to real data. In this study, it is aimed to compare the performances of SMOTE, SMOTEENN, BorderlineSMOTE, SMOTETomek and ADASYN methods that have been used in synthetic data production by considering the importance of synthetic data generation in line with the increasing need for data use in the health field. In the study, a dataset consisting of 15 variables belonging to 390 patients with different observation and class numbers and a dataset consisting of 16 variables related to 19,212 COVID-19 patients were used. It has been concluded that SMOTE is more successful in balancing the data sets with large number of observations and multiclass classification. This technique can be used effectively in synthetic data generation compared to hybrid techniques.