Wimbush, Alexander
(2023)
Propagation of Epistemic Uncertainty Through Medical Diagnostic Algorithms.
PhD thesis, University of Liverpool.
|
Text
200646415_Aug2023.pdf - Author Accepted Manuscript Download (35MB) | Preview |
Abstract
Medical diagnostic algorithms combine multiple tests to improve medical decision-making. These are commonly developed case by case with assessment of tests, or through data-driven approaches. This thesis presents a means of assessing diagnostic algorithm characteristics under aleatory uncertainty so that they may be combined in a genetic programming optimiser to identify suitable combinations of tests when test characteristics are known. When a lack of data prevents the performance of individual tests, or the performance of groups of tests, being known precisely, this uncertainty can be characterised as epistemic uncertainty. Confidence curves are used to characterise and propagate epistemic uncertainty using a simple method which demonstrably maintains well-calibrated statistical coverage with no dependence assumptions. The characteristics of the statistical coverage of these methods is assessed visually using plots of coverage rates against confidence levels. These plots are used to demonstrate the utility of confidence curves developed for the negative binomial case, along with maximum-sample negative binomial and Poisson distributions. Confidence curves demonstrably provide much tighter centred intervals than equivalent intervals taken from confidence boxes while maintaining strict statistical coverage under arbitrary dependence. If confidence curves are not available, this thesis presents a method by which they can be reconstructed to facilitate easy propagation. The final contribution demonstrates a method for assessing test characteristics when the true states of the test population cannot be precisely known. This is analogous to latent class modelling using confidence curves rather than probability distributions, and shows good statistical coverage as a result. Together, these tools allow for characterisation and propagation of epistemic uncertainty in a manner that facilitates medical decision-making in low-data environments.
| Item Type: | Thesis (PhD) |
|---|---|
| Divisions: | Faculty of Science and Engineering > School of Engineering |
| Depositing User: | Symplectic Admin |
| Date Deposited: | 30 Aug 2023 09:21 |
| Last Modified: | 23 Jan 2024 11:20 |
| DOI: | 10.17638/03172225 |
| Supervisors: |
|
| URI: | https://livrepository.liverpool.ac.uk/id/eprint/3172225 |
Dimensions
Dimensions
