High-dimensional limited-sample biomedical data classification using variational autoencoder

Abstract

Biomedical prediction is vital to the modern scientific view of life, but it is a challenging task due to high-dimensionality, limited-sample size (also known as HDLSS problem), non-linearity, and data types tend are complex. A large number of dimensionality reduction techniques developed, but, unfortunately, not efficient with small-sample (observation) size dataset. To overcome the pitfalls of the sample-size and dimensionality this study employed variational autoencoder (VAE), which is a powerful framework for unsupervised learning in recent years. The aim of this study is to investigate a reliable biomedical diagnosis method for HDLSS dataset with minimal error. Hence, to evaluate the strength of the proposed model six genomic microarray datasets from Kent Ridge Repository were applied. In the experiment, several choices of dimensions were selected for data preprocessing. Moreover, to find a stable and suitable classifier, different popular classifiers were applied. The experimental results found that the VAE can provide superior performance compared to the traditional methods such as PCA, fastICA, FA, NMF, and LDA.