Use this URL to cite or link to this record in EThOS:
Title: Bayesian fusion of continuous-valued labels in biomedical applications
Author: Zhu, Tingting
ISNI:       0000 0004 6498 0140
Awarding Body: University of Oxford
Current Institution: University of Oxford
Date of Award: 2016
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Expert labelling is the gold standard for diagnosing patient-specific diseases from medical data. However, experts are relatively scarce, their time is expensive, and the task of labelling is time-consuming. While automated algorithms offer advantages of time efficiency, repeatability, and cost-saving benefits compared to manual labelling, there remains a substantial discrepancy in their estimation as well as reliability that limit their use. Two commonly-encountered and clinically-important examples were considered for the use of concept labels in clinical practice: the QT interval in the electrocardiogram and the respiratory rate estimation from the photoplethysmogram. This thesis sought to combine outputs of automated annotators (i.e., algorithms) to improve estimation of these exemplars in a principled manner. A series of methods were proposed for inferring a consensus in a scenario when only noisy annotations of some presumed underlying ground truth are provided: (1) the Bayesian Continuous-Valued Label Aggregator (BCLA) with independent annotators was proposed to infer the ground truth while jointly predicting each algorithm's bias and precision. Both maximum-a-posteriori and Gibbs sampling approaches were derived to compute parameters in the BCLA model (denoted as BCLA-MAP and BCLA-Gibbs, respectively); (2) BCLA incorporating the notion of signal quality (i.e., BCLAs) was also proposed to provide a confidence measure relating to the quality of the input recordings; (3) BCLA with correlated annotators (BCLAc) model was further proposed to provide a more generalised framework for BCLA, in particular, two variants of BCLAc were considered to model the correlation of errors among annotators directly and indirectly. Furthermore, the proposed methods were compared to the mean, median, best-performing single annotator with least root-meansquared- error, and two state-of-the-art benchmarking approaches described in the literature, namely "scalar Simultaneous Truth and Performance Level Estimation" and an Expectation- Maximisation label aggregation approach. This thesis considered two exemplary medical datasets as well as simulated datasets to validate the proposed methods. The results of the proposed BCLA had optimal performance over the other comparison voting strategies in all datasets. Furthermore, BCLA-Gibbs as a fully-Bayesian approach was superior to BCLA-MAP. When incorporating an extension to include signal quality (i.e., the BCLAs model), results of BCLAs-MAP and BCLAs-Gibbs outperformed those of the basic BCLA methods consistently. Finally, we modelled the inclusion of potentially-correlated annotators using BCLAc. In summary, the results of the methods proposed in this thesis improve accuracy by inferring a consensus in a scenario when only noisy labels of a group of imperfect algorithms are available. The proposed methods operate in an unsupervised Bayesian learning framework, and have the potential for real-time application to produce estimates that are more robust than any of the algorithms independently considered.
Supervisor: Clifton, David A. ; Clifford, Gari D. Sponsor: RCUK Digital Economy Programme
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available