Abstract

Analysis of medical data and making precise decisions by machine learning is emerging as a hot topic in healthcare. The ultimate goal of using these techniques is to transform data into actionable knowledge for supporting clinicians and improving patients quality of life. To assist health professionals in making precise decisions, clustering is among the most applied methods. This approach aims to stratify patients into meaningful groups based on their similarities in medical data spaces such as images, signals, and medical records. Among all clustering approaches, mixture models have been widely used by researchers in different fields due to their substantial flexibility to explain the data. Traditionally, Gaussian mixture models have been applied in real-world applications, but data are not Gaussian in many fields. In this thesis, we proposed novel clustering approaches based on finite and infinite mixture models. Our mixture models are developed based on a new distribution called multivariate Beta distribution which demonstrated lots of flexibility to fit data of different shapes. We paired our models with capable learning methods such as variational inference and expectation propagation. These learning methods determine the correct number of mixture components and estimate model's parameters which are two known challenges while fitting mixture models. Moreover, we modeled sequential data and developed a novel version of the hidden Markov model, namely multivariate Beta-based hidden Markov model, and extended it to a nonparametric model to increase its flexibility. All developed models are evaluated on real medical applications including medical images and signals. The outcomes demonstrated that our proposed models outperform similar alternatives.