New methods for algorithm evaluation and cluster initialisation with applications to healthcare

Wilde, Henry David 2021. New methods for algorithm evaluation and cluster initialisation with applications to healthcare. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF - Accepted Post-Print Version Download (5MB) | Preview
	PDF (Cardiff University Electronic Publication Form.) - Supplemental Material Restricted to Repository staff only Download (248kB)

Abstract

This thesis explores three themes related to modern operational research: evaluating the objective performance of an algorithm, combining clustering with concepts of mathematical fairness, and developing insightful healthcare models despite a lack of fine-grained data. The established evaluation procedure for algorithms — and particularly machine learning algorithms — lacks robustness, potentially inflating the success of the methods being assessed. To tackle this, the evolutionary dataset optimisation method is introduced as a supplementary evaluation tool. By traversing the space in which datasets exist, this method provides the means of attaining a richer understanding of the algorithm under study. This method is used to investigate a novel initialisation method for a centroid-based clustering algorithm, k-modes. The initialisation makes use of the game theoretic concept of a matching game to allocate the starting centroids in a mathematically fair way. The subsequent investigation reveals the conditions under which the new initialisation improves upon two other initialisation methods. An extension to the k-modes algorithmis utilised to segment an administrative dataset provided by the co-sponsors of this project, CwmTaf MorgannwgUniversity Health Board. The dataset corresponds to the patient population presenting a specific chronic disease, and comprises a high-level summary of their stays in hospital over a number of years. Despite the relative coarseness of this dataset, the segmentation provides a useful profiling of its instances. These profiles are used to inform a multi-class queuing model representing a hypothetical ward for the affected patients. Following a novel validation process for the queuing model, actionable insights into the needs of the population are found. In addition to these research pursuits, several open-source software packages have been developed to accompany this thesis. These pieces of software were developed using best practices to ensure the reliability, reproducibility, and sustainability of the research in this thesis

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Mathematics
Subjects:	Q Science > QA Mathematics
Funders:	Cwm Taf Morgannwg University Health Board
Date of First Compliant Deposit:	15 April 2021
Last Modified:	27 Sep 2022 01:09
URI:	https://orca.cardiff.ac.uk/id/eprint/140492

Actions (repository staff only)

Edit Item