Machine Learning Models for Multidimensional Clinical Data

Abstract

Healthcare monitoring systems in the hospital and at home generate large quantities of rich-phenotype data from a wide array of sources. Typical sources include clinical observations, continuous waveforms, lab results, medical images and text notes. The key clinical challenge is to interpret these in a way that helps to improve the standard of patient care. However, the size and complexity of the data sets, which are often multidimensional and dynamically changing, means that interpretation is extremely difficult, even for expert clinicians. One important set of approaches to this challenge is Machine Learning Systems. These are systems that analyse and interpret data in a way that automatically recognizes underlying patterns and trends. These patterns are useful for predicting future clinical events such as hospital re-admission, and for determining rules within clinical decision support tools. In this chapter we will provide a review of machine learning models currently used for event prediction and decision support in healthcare monitoring. In particular, we highlight how these approaches deal with multi-dimensional data. We then discuss some of the practical problems in implementing Machine Learning Systems. These include: missing or corrupted data, incorporation of heterogeneous and multimodal data, and generalization across patient populations and clinical settings. Finally, we discuss promising future research directions, including the most recent developments in Deep Learning.