Predictive analytics in an intensive care unit by processing streams of physiological data in real-time

Abstract

Computing systems deployed in hospital environments routinely collect a large volume of data that has not, thus far, been widely examined. There has been increasing research into the application of machine learning techniques on these data streams to predict and prevent disease states, however limitations exist around understanding the optimal methods and availability of data to train such models which is what we aim to address in this thesis. This thesis provides an insight into the work that has already been carried out in this field and further presents state-of-the-art research of AI for prediction of physiological data across two problem domains. This thesis features a description of the first Northern Ireland research-based ICU database that will enable further research to advance the field, of which we have set up. We compare machine learning methods for the classification of cardiovascular disease. By exploring support vector machines, artificial neural networks and four ensemble methods, based on decision trees, the results show how varying their hyperparameters can affect the accuracy of the predictions. Further, this work uses two public datasets with differing characteristics in order to understand the potential differences in the uncertainty of the methods. This thesis further presents the first application of machine learning to enable lung protective ventilation. We successfully predict a given ventilator parameter within 10% accuracy of its true value and can in turn predict an average of 70% of alerts per patient, meaning clinicians can intervene and prevent injury from occurring.

Date of Award	Dec 2022
Original language	English
Awarding Institution	Queen's University Belfast
Sponsors	Northern Ireland Department for the Economy
Supervisor	Charles Gillan (Supervisor), Ivor Spence (Supervisor) & Murali Shyamsundar (Supervisor)