Combining Synthetic Patient Data Generation with Machine Learning Methods for Diabetes Prediction

Abstract

Diabetes Mellitus (DM) is a chronic disease caused by different disorders in the insulin production or use. Its prevalence has not stopped increasing during the last years, becoming a major public health concern. Thus, tools for its prediction and early diagnosis are needed. In this context, Machine Learning (ML) could be a suitable choice due to its capability of extracting useful information from medical records. However, the lack of available and reliable datasets makes this a complex task. Synthetic data generation is emerging as a solution for this issue, as it takes a real dataset as the basis to generate similar instances. In this work, a framework based on ML and synthetic data generation methods is presented to evaluate whether classification performance between presence or absence of DM could be improved. The obtained results show that ADASYN and Borderline SMOTE algorithms fairly keep the underlying structure of the original data. They also prove that the ML models trained with mixed synthetic and original data perform as well as those trained with original data.