Entropy-Based Variational Inference for Semi-Bounded Data Clustering in Medical Applications

Abstract

Over the past decades, the unprecedented availability of various types of data and simultaneous development of technology established extensive interest in applying numerous machine learning approaches to extract the implicit patterns, acquire information and retrieve latent meaningful knowledge. Such powerful statistical tools have been applied in various fields of science.

One of the vital domains where these techniques could be potentially deployed is healthcare. The main intention in this field is that medical diagnosis procedures and healthcare examinations generate a huge amount of various data types such as text, image, video and signal. Dealing with such large complex data is beyond the scope of human competence. Consequently, machine learning tools are significantly valuable as they assist the clinicians in processing medical datasets, achieving broader insight, planning and managing diseases, providing better care which leads to having better outcomes including elimination of unnecessary costs and increasing patient satisfaction.

In this work, we focus on one of the main clustering methods of machine learning approaches, namely mixture models. These capable techniques have demonstrated high potential and flexibility to express data. Gaussian mixture models (GMM) have been widely applied in various fields of research to express symmetric data. However, for asymmetric and non-Gaussian data, other alternatives such as inverted Dirichlet mixture models could describe the data more accurately. To learn our model, we employ an entropy-based variational approach and then evaluate it on four medical applications.