| dc.description.abstract | Clinical decision support is needed for chronic medical conditions. Compared with a decision about the treatment choice in an acute condition, the management of a chronic condition involves repeated and regular decisions. Moreover, these decisions can be shared between the patient and the clinicians, allowing the patient some degree of self-management. Patient entered data also enables self-management but, without some degree of automated decision-support, may raise the clinical workload unacceptably. This thesis addresses the problem of decision support in the management of chronic medical conditions using Bayesian networks (BNs). BNs have been widely proposed for clinical decision support, but only a few published studies address the management of chronic disease and there is scarce evidence of their adoption into practice. Their advantages include their graphical structure which makes it possible to build a BN from experts’ knowledge. This is particularly valuable when data is scarce, which is to be expected when we try to adopt new patterns of clinical management. In this thesis, we show how BNs can be used to assist screening, diagnosis, care allocation and self-management of chronic conditions using a case study in Gestational Diabetes Mellitus (GDM). This work is based on the collaboration with clinical team composed by a Professor in diabetes, a midwife and two clinicians working with GDM in Royal London Hospital (RLH) and Newham Hospital (NH). GDM is a condition that affects a large number of pregnant women in the UK and, if not managed appropriately, can lead to adverse outcomes for both mother and child. It can be diagnosed at different stages in pregnancy and the timing and choice of diagnostic tests depends on the presence of risk factors and previous test results. Diagnosis of GDM is primarily based on blood glucose test results but the best thresholds for diagnosis and treatment, which ranges from diet advice to medication, are not clear and the cost-effectiveness of managing GDM has been questioned. The underlying issue is that there is no ‘gold-standard’ diagnostic test for GDM that could be used to optimise the thresholds of the tests used in practice. To resolve this problem of diagnostic uncertainty, we propose a BN to predict the final outcome in terms of the treatment, if any, used to control the blood glucose of pregnant women. This BN can be used to assist decisions about screening and diagnosis, such as discharging women from further testing at different stages in pregnancy to avoid unnecessary clinic visits. Once diagnosed with GDM, pregnant women are seen in the midwifery clinic or hospital depending on the treatment prescribed and obstetric risks. We believe this BN can be also used to identify pregnant women who are more likely to require intensive care and medication to control their BGL and inform decisions regarding the care pathway. The structure of the BN was built from experts’ knowledge and two datasets were used to obtain the BN parameters. We illustrate how the model could be used to inform the screening, diagnostic strategies and care allocation decisions using real data. People with GDM need to follow a diet and exercise routine and, occasionally, take medication as part of the GDM treatment to control their blood glucose levels. Research supports that blood glucose control is crucial to reduce the risk of adverse outcomes and frequent blood glucose monitoring is advised as part of GDM management routine. GDM treatment is adjusted during clinic or hospital visits that, in general, happen fortnightly and can be demanding for both pregnant women and clinicians. People not taking medication and whose BGL is under control would benefit from a system that could offer advice as to when an appointment should be scheduled, potentially avoiding unnecessary clinic/hospital visits. We proposed a Dynamic Bayesian Network (DBN) build from expert’s knowledge that can assist people with GDM to manage glucose control when medication is not being used. We exemplify how the model can be used and test the model’s overall ability to suggest an appointment in the correct time using data from 117 people diagnosed with GDM. For some people, it is harder to manage their BGL without medication either because they are unable to adhere to a diet and exercise routine or because their blood glucose control deteriorates as a result of GDM. Hospital appointments are scheduled to revise the treatment and adjust the medication dose. If the blood glucose levels are not controlled after the medication adjustment, people wait until the next appointment for further adjustments to the treatment. However, if the prescription successfully controls the blood glucose levels, people still attend the appointment even though no dose change is necessary. Again using expert knowledge, we propose an extension to the DBN to offer advice to people on how to adjust their medication dose and when to schedule an appointment. We illustrate how the model works with real data. Many AI (or ML) systems have been proposed for clinical decision support. Clinical usefulness is assessed in an ‘Impact Study’, a form of trial of a completed system. In development, in contrast, the focus is on AI accuracy measures, such as the AUC. As prediction accuracy does not ensure clinical utility, to improve the impact and to justify the cost of a study, the potential impact of a proposed AI system should be modelled during its development. We show that an Influence Diagram (ID) can be used for this and provide a small set of generic models for diagnostic AI systems. We show that how the AI interacts with clinical decision makers is at least as important as its predictive accuracy. We use information about the models developed in this thesis and GDM to illustrate the use of an Influence Diagram to assess the potential impact of systems based on these models. | en_US |
