Year: 2016
Type: dissertation
Abstract: This dissertation reports on a series of five studies that were motivated by the need to incorporate information about similar patients into clinical decision making. The studies examined questions relating to the feasibility of a tool that could identify similar patients in health data repositories, and then use the resulting patient types for prediction of unknown values in the health record for a current patient. The main question addressed in this dissertation was: how well do predictions based on summarized clusters match predictions based on the original (unsummarized) data in terms of accuracy? The motivation for using summarized prediction in this dissertation is that summarization of sufficiently large numbers of cases removes the privacy concerns typically associated with using individual data. After an introductory chapter which establishes the motivation for the research and the overall research strategy, each of the five following chapters presents a case study in the use of cluster analysis in secondary analysis of patient data. Chapters 2, 3 and 4 utilized (MIMIC II) intensive care data. A set of patient types was developed in chapter 2 and the accuracy of predictions made with summarized data based on patient types was assessed in chapter 3. Chapter 4 focuses on the role of feature extraction in clinical prediction and compares the method used in chapter 3 with some other methods such as the K-nearest Neighbor method. Chapters 5 and 6 utilized the (ICES) cancer data from the province of Ontario. A set of patient types was developed from the cancer data (in chapter 5) and in chapter 6 the accuracy of predictions made with summarized data, based on the set of cancer patient types, was evaluated. The dissertation concludes in the final chapter with a review of the contributions (both empirical and methodological) made in the research as well as a listing of the limitations of the research and future research prospects. The main finding of this research was that relatively accurate predictions can be made based on summarized patient data in a way that does not raise privacy concerns. A secondary result is that when accuracy in prediction is used to select a particular partition of the patients, the resulted patient types can be quite interpretable. The present results point the way towards improved clinical decision support tools, and, possibly, the development of generalized health data search engines.
Source: TSpace (University of Toronto)
Author: Mahsa Rouzbahman
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2626129713