Application of a New VAE-MF Generative Model in TCD Dataset

Abstract

Transcranial Doppler (TCD) is a non-invasive method for detecting ischemic stroke disease and is widely used in clinical diagnosis. Due to the obvious imbalance of medical data, oversampling is needed to balance it. Although the oversampling method is one of the important means to solve the problem of imbalanced data classification, the traditional oversampling method will inevitably introduce noise, which will affect the classification results. To address this issue, we proposed a new oversampling method combining the membership function (MF) and the variational autoencoder (VAE). Our method utilized VAE as a generative model to generate new samples to reduce the Imbalance Ratio (IR) of the dataset. Then the new samples are filtered using MF to weaken the impact of the noise introduced by the oversampling method on classification performance. In addition, in view of the weak adaptability of traditional MFs to complex sample distributions in practical applications, a new MF is proposed by using kernel function mapping and hypersphere. The classification experiments on TCD imbalanced dataset prove that the oversampling method we proposed has improved classification performance on multiple evaluation criteria compared with traditional oversampling methods.