dc.description.abstract | Bayesian networks (BNs) have been widely proposed for medical decision support. One
advantage of a BN is reasoning under uncertainty, which is pervasive in medicine. Another
advantage is that a BN can be built from both data and knowledge and so can be applied in
circumstances where a complete dataset is not available. In this thesis, we examine how BNs
can be used for the decision support challenges of chronic diseases. As a case study, we study
Rheumatoid Arthritis (RA), which is a chronic inflammatory disease causing swollen and
painful joints. The work has been done as part of a collaborative project including clinicians
from Barts and the London NHS Trust involved in the treatment of RA. The work covers
three stages of decision support, with progressively less available data.
The first decision support stage is diagnosis. Various criteria have been proposed by
clinicians for early diagnosis but these criteria are deterministic and so do not capture
diagnostic uncertainty, which is a concern for patients with mild symptoms in the early
stages of the disease. We address this problem by building a BN model for diagnosing
RA. The diagnostic BN model is built using both a dataset of 360 patients provided by the
clinicians and their knowledge as experts in this domain. The choice of factors to include
in the diagnostic model is informed by knowledge, including a model of the care pathway
which shows what information is available for diagnosis. Knowledge is used to classify the
factors as risk factors, relevant comorbidities, evidence of pathogenesis mechanism, signs,
symptoms, and serology results, so that the structure of BN model matches the clinical
understanding of RA.
Since most of the factors are present in the dataset, we are able to train the parameters
of the diagnostic BN from the data. This diagnostic BN model obtains promising results
in differentiating RA cases from other inflammatory arthritis cases. Aware that eliciting
knowledge is time-consuming and could limit the uptake of these techniques, we consider
two alternative approaches. First, we compare its diagnostic performance with an alternative
BN model entirely learnt from data; we argue that having a clinically meaningful structure
allows us to explain clinical scenarios in a way that cannot be done with the model learnt
purely from data. We also examine whether useful knowledge can be retrieved from existing
vi
medical ontologies, such as SNOMED CT and UMLS. Preliminary results show that it could
be feasible to use such sources to partially automate knowledge collection.
After patients have been diagnosed with RA, they are monitored regularly by a clinical
team until the activity of their disease becomes low. The typical care arrangement has two
challenges: first, regular meetings with clinicians occur infrequently at fixed intervals (e.g.,
every six months), during which time the activity of the disease can increase (or ‘flare’) and
decrease several times. Secondly, the best medications or combinations of medications must
be found for each patient, but changes can only be made when the patient visits the clinic. We
therefore develop this stage of decision support in two parts: the first and simplest part looks
at how the frequency of clinic appointments could be varied; the second part builds on this to
support decisions to adjust medication dosage. We describe this as the ‘self-management’
decision support model.
Disease activity is commonly measured with Disease Activity Score 28 (DAS28). Since
the joint count parts of this can be assessed by the patient, the possibility of collecting regular
(e.g., weekly) DAS28 data has been proposed. It is not yet in wide use, perhaps because of
the overheads to the clinical team of reviewing data regularly. The dataset available to us
for this work came from a feasibility study conducted by the clinical collaborators of one
system for collecting data from patients, although the frequency is only quarterly. The aim of
the ‘self-management’ decision support system is therefore to sit between patient-entered
data and the clinical team, saving the work of clinically assessing all the data. Specifically,
in the first part we wish to predict disease activity so that an appointment should be made
sooner, distinguishing this from patients whose disease is well-managed so that the interval
between appointments can be increased. To achieve this, we build a dynamic BN (DBN)
model to monitor disease activity and to indicate to patients and their clinicians whether a
clinical review is needed. We use the data and a set of dummy patient scenarios designed by
the experts to evaluate the performance of the DBN.
The second part of the ‘self-management’ decision support stage extends the DBN to
give advice on adjustments to the medication dosage. This is of particular clinical interest
since one class of medications used (biological disease-modifying antirheumatic drugs) are
very expensive and, although effective at reducing disease activity, can have severe adverse
reactions. For both these reasons, decision support that allowed a patient to ‘taper’ the dosage
of medications without frequent clinic visits would be very useful. This extension does not
meet all the decision support needs, which ideally would also cover decision-making about
the choice of medications. However, we have found that as yet there is neither sufficient data
nor knowledge for this.
vii
The third and final stage of decision support is targeted at patients who live with RA. RA
can have profound impacts on the quality of life (QoL) of those who live with it, affecting
work, financial status, friendships, and relationships. Information from patient organisations
such as the leaflets prepared by the National Rheumatoid Arthritis Society (NRAS) contains
advice on managing QoL, but the advice is generic, leaving it up to each patient to select the
advice most relevant to their specific circumstances. Our aim is therefore to build a BN-based
decision support system to personalise the recommendations for enhancing the QoL of RA
patients. We have built a BN to infer three components of QoL (independence, participation,
and empowerment) and shown how this can be used to target advice. Since there is no
data, the BN is developed from expert knowledge and literature. To evaluate the resulting
system, including the BN, we use a set of patient interviews conducted and coded by our
collaborators. The recommendations of the system were compared with those of experts in a
set of test scenarios created from the interviews; the comparison shows promising results. | en_US |