Iñigo Bermejo

Decision analysis networks

Abstract This paper presents decision analysis networks (DANs) as a new type of probabilistic graphical model. Like influence diagrams (IDs), DANs are much more compact and easier to build than decision trees and can represent conditional independencies. In fact, for every ID there is an equivalent symmetric DAN, but DANs can also represent asymmetric problems involving partial orderings of the decisions (order asymmetry), restrictions between the values of the variables (domain asymmetry), and conditional observability (information asymmetry). Symmetric DANs can be evaluated with the same algorithms as IDs. Every asymmetric DAN can be evaluated by converting it into an equivalent decision tree or, much more efficiently, by decomposing it into a tree of symmetric DANs. Given that DANs can solve symmetric problems as easily and as efficiently as IDs, and are more appropriate for asymmetric problems—which include virtually all real-world problems—DANs might replace IDs as the standard type of probabilistic graphical model for decision support and decision analysis. We also argue that DANs compare favorably with other formalisms proposed for asymmetric decision problems. In practice, DANs can be built and evaluated with OpenMarkov, a Java open-source package for probabilistic graphical models.

Markov Influence Diagrams.

Markov influence diagrams (MIDs) are a new type of probabilistic graphical model that extends influence diagrams in the same way that Markov decision trees extend decision trees. They have been designed to build state-transition models, mainly in medicine, and perform cost-effectiveness analyses. Using a causal graph that may contain several variables per cycle, MIDs can model various patient characteristics without multiplying the number of states; in particular, they can represent the history of the patient without using tunnel states. OpenMarkov, an open-source tool, allows the decision analyst to build and evaluate MIDs—including cost-effectiveness analysis and several types of deterministic and probabilistic sensitivity analysis—with a graphical user interface, without writing any code. This way, MIDs can be used to easily build and evaluate complex models whose implementation as spreadsheets or decision trees would be cumbersome or unfeasible in practice. Furthermore, many problems that previously required discrete event simulation can be solved with MIDs; i.e., within the paradigm of state-transition models, in which many health economists feel more comfortable.

Replicating Health Economic Models: Firm Foundations or a House of Cards?

Health economic evaluation is a framework for the comparative analysis of the incremental health gains and costs associated with competing decision alternatives. The process of developing health economic models is usually complex, financially expensive and time-consuming. For these reasons, model development is sometimes based on previous model-based analyses; this endeavour is usually referred to as model replication. Such model replication activity may involve the comprehensive reproduction of an existing model or ‘borrowing’ all or part of a previously developed model structure. Generally speaking, the replication of an existing model may require substantially less effort than developing a new de novo model by bypassing, or undertaking in only a perfunctory manner, certain aspects of model development such as the development of a complete conceptual model and/or comprehensive literature searching for model parameters. A further motivation for model replication may be to draw on the credibility or prestige of previous analyses that have been published and/or used to inform decision making. The acceptability and appropriateness of replicating models depends on the decision-making context: there exists a trade-off between the ‘savings’ afforded by model replication and the potential ‘costs’ associated with reduced model credibility due to the omission of certain stages of model development. This paper provides an overview of the different levels of, and motivations for, replicating health economic models, and discusses the advantages, disadvantages and caveats associated with this type of modelling activity. Irrespective of whether replicated models should be considered appropriate or not, complete replicability is generally accepted as a desirable property of health economic models, as reflected in critical appraisal checklists and good practice guidelines. To this end, the feasibility of comprehensive model replication is explored empirically across a small number of recent case studies. Recommendations are put forward for improving reporting standards to enhance comprehensive model replicability.

Computer-assisted CI fitting: Is the learning capacity of the intelligent agent FOX beneficial for speech understanding?

Objective: The software application FOX (‘Fitting to Outcome eXpert’) is an intelligent agent to assist in the programing of cochlear implant (CI) processors. The current version utilizes a mixture of deterministic and probabilistic logic which is able to improve over time through a learning effect. This study aimed at assessing whether this learning capacity yields measurable improvements in speech understanding. Methods: A retrospective study was performed on 25 consecutive CI recipients with a median CI use experience of 10 years who came for their annual CI follow-up fitting session. All subjects were assessed by means of speech audiometry with open set monosyllables at 40, 55, 70, and 85 dB SPL in quiet with their home MAP. Other psychoacoustic tests were executed depending on the audiologists clinical judgment. The home MAP and the corresponding test results were entered into FOX. If FOX suggested to make MAP changes, they were implemented and another speech audiometry was performed with the new MAP. Results: FOX suggested MAP changes in 21 subjects (84%). The within-subject comparison showed a significant median improvement of 10, 3, 1, and 7% at 40, 55, 70, and 85 dB SPL, respectively. All but two subjects showed an instantaneous improvement in their mean speech audiometric score. Discussion: Persons with long-term CI use, who received a FOX-assisted CI fitting at least 6 months ago, display improved speech understanding after MAP modifications, as recommended by the current version of FOX. This can be explained only by intrinsic improvements in FOXs algorithms, as they have resulted from learning. This learning is an inherent feature of artificial intelligence and it may yield measurable benefit in speech understanding even in long-term CI recipients.

A Probabilistic Graphical Model for Tuning Cochlear Implants

Severe and profound hearing losses can be treated with cochlear implants (CI). Given that a CI may have up to 150 tunable parameters, adjusting them is a highly complex task. For this reason, we decided to build a decision support system based on a new type of probabilistic graphical model (PGM) that we call tuning networks. Given the results of a set of audiological tests and the current status of the parameter set, the system looks for the set of changes in the parameters of the CI that will lead to the biggest improvement in the user’s hearing ability. Because of the high number of variables involved in the problem we have used an object-oriented approach to build the network. The prototype has been informally evaluated comparing its advice with those of the expert and of a previous decision support system based on deterministic rules. Tuning networks can be used to adjust other electrical or mechanical devices, not only in medicine.

Advanced Algorithms for Medical Decision Analysis. Implementation in OpenMarkov

In spite the important advantages of influence diagrams over decision trees, including the possibility of solving much more complex problems, the medical literature still contains around 10 decision trees for each influence diagram. In this paper we analyse the reasons for the low acceptance of influence diagrams in health decision analysis, in contrast with its success in artificial intelligence. One of the reasons is the difficulty of representing asymmetric problems. Another one was the lack of algorithms for explaining the reasoning and performing cost-effectiveness analysis, as well as the scarcity of user-friendly software tools for sensitivity analysis. In this paper we review the research conducted by our group in the last 25 years, crystallised in the open-source software tool OpenMarkov, explaining how it has tried to address those challenges.

Response to ‘Comment on “Replicating Health Economic Models: Firm Foundations or a House of Cards?”’

eprints@whiterose.ac.uk https://eprints.whiterose.ac.uk/ Reuse Items deposited in White Rose Research Online are protected by copyright, with all rights reserved unless indicated otherwise. They may be downloaded and/or printed for private study, or other acts as permitted by national copyright laws. The publisher or other rights holders may allow further reproduction and re-use of the full text version. This is indicated by the licence information on the White Rose Research Online record for the item.

Certolizumab Pegol for Treating Rheumatoid Arthritis Following Inadequate Response to a TNF-α Inhibitor: An Evidence Review Group Perspective of a NICE Single Technology Appraisal

As part of its single technology appraisal (STA) process, the National Institute for Health and Care Excellence (NICE) invited the manufacturer (UCB Pharma) of certolizumab pegol (CZP; Cimzia®) to submit evidence of its clinical and cost effectiveness for the treatment of rheumatoid arthritis (RA) following inadequate response to a tumour necrosis factor-α inhibitor (TNFi). The School of Health and Related Research Technology Appraisal Group at the University of Sheffield was commissioned to act as the independent Evidence Review Group (ERG). The ERG produced a detailed review of the evidence for the clinical and cost effectiveness of the technology, based upon the company’s submission to NICE. The clinical effectiveness evidence in the company’s submission for CZP was based predominantly on six randomised controlled trials (RCTs) comparing the efficacy of CZP against placebo. The clinical effectiveness review identified no head-to-head evidence on the efficacy of CZP against the comparators within the scope; therefore, the company performed a network meta-analysis (NMA). The company’s NMA concluded that CZP had a similar efficacy to that of its comparators. The company submitted a Markov model that assessed the incremental cost effectiveness of CZP versus comparator biologic disease-modifying antirheumatic drugs (bDMARDs) for the treatment of RA from the perspective of the National Health Service for three decision problems, each of which followed an inadequate response to a TNFi. These were (1) a comparison against rituximab (RTX) in combination with methotrexate (MTX); (2) a comparison against bDMARDs when RTX was contraindicated or withdrawn due to an adverse event; and (3) a comparison against bDMARDs when MTX was contraindicated or withdrawn due to an adverse event. Results from the company’s economic evaluation showed that CZP resulted in a similar number of quality-adjusted life years (QALYs) produced at similar or lower costs compared with comparator bDMARDs. The commercial-in-confidence patient access schemes for abatacept and tocilizumab could not be incorporated by the company, but were incorporated by the ERG in a confidential appendix for the NICE Appraisal Committee (AC). The company estimated that the addition of CZP before RTX in a sequence for patients who could receive MTX produced more QALYs at an increased cost, with a cost per QALY of £33,222. Following a critique of the model, the ERG undertook exploratory analyses that did not change the conclusions reached based on the company’s economic evaluation in relation to the comparison with bDMARDs. The ERG estimated that where CZP replaced RTX, CZP was dominated, as it produced fewer QALYs at an increased cost. The AC concluded that there was little difference in effectiveness between CZP and comparator bDMARDs and that equivalence among bDMARDs could be accepted. The AC consequently recommended CZP plus MTX for people for whom RTX is contraindicated or not tolerated and CZP monotherapy for people for whom MTX is contraindicated or not tolerated. The AC concluded that CZP plus MTX could not be considered a cost-effective use of National Health Service resources when RTX plus MTX is a treatment option.

A systematic review and economic evaluation of adalimumab and dexamethasone for treating non-infectious intermediate uveitis, posterior uveitis or panuveitis in adults

BACKGROUND Non-infectious intermediate uveitis, posterior uveitis and panuveitis are a heterogeneous group of inflammatory eye disorders. Management includes local and systemic corticosteroids, immunosuppressants and biological drugs. OBJECTIVES To evaluate the clinical effectiveness and cost-effectiveness of subcutaneous adalimumab (Humira®; AbbVie Ltd, Maidenhead, UK) and a dexamethasone intravitreal implant (Ozurdex®; Allergan Ltd, Marlow, UK) in adults with non-infectious intermediate uveitis, posterior uveitis or panuveitis. DATA SOURCES Electronic databases and clinical trials registries including MEDLINE, EMBASE, Cochrane Database of Systematic Reviews, Database of Abstracts of Reviews of Effects and the World Health Organizations International Clinical Trials Registry Platform were searched to June 2016, with an update search carried out in October 2016. REVIEW METHODS Review methods followed published guidelines. A Markov model was developed to assess the cost-effectiveness of dexamethasone and adalimumab, each compared with current practice, from a NHS and Personal Social Services (PSS) perspective over a lifetime horizon, parameterised with published evidence. Costs and benefits were discounted at 3.5%. Substantial sensitivity analyses were undertaken. RESULTS Of the 134 full-text articles screened, three studies (four articles) were included in the clinical effectiveness review. Two randomised controlled trials (RCTs) [VISUAL I (active uveitis) and VISUAL II (inactive uveitis)] compared adalimumab with placebo, with limited standard care also provided in both arms. Time to treatment failure (reduced visual acuity, intraocular inflammation, new vascular lesions) was longer in the adalimumab group than in the placebo group, with a hazard ratio of 0.50 [95% confidence interval (CI) 0.36 to 0.70; p < 0.001] in the VISUAL I trial and 0.57 (95% CI 0.39 to 0.84; p = 0.004) in the VISUAL II trial. The adalimumab group showed a significantly greater improvement than the placebo group in the 25-item Visual Function Questionnaire (VFQ-25) composite score in the VISUAL I trial (mean difference 4.20; p = 0.010) but not the VISUAL II trial (mean difference 2.12; p = 0.16). Some systemic adverse effects occurred more frequently with adalimumab than with placebo. One RCT [HURON (active uveitis)] compared a single 0.7-mg dexamethasone implant against a sham procedure, with limited standard care also provided in both arms. Dexamethasone provided significant benefits over the sham procedure at 8 and 26 weeks in the percentage of patients with a vitreous haze score of zero (p < 0.014), the mean best corrected visual acuity improvement (p ≤ 0.002) and the percentage of patients with a ≥ 5-point improvement in VFQ-25 score (p < 0.05). Raised intraocular pressure and cataracts occurred more frequently with dexamethasone than with the sham procedure. The incremental cost-effectiveness ratio (ICER) for one dexamethasone implant in one eye for a combination of patients with unilateral and bilateral uveitis compared with limited current practice, as per the HURON trial, was estimated to be £19,509 per quality-adjusted life-year (QALY) gained. The ICER of adalimumab for patients with mainly bilateral uveitis compared with limited current practice, as per the VISUAL trials, was estimated to be £94,523 and £317,547 per QALY gained in active and inactive uveitis respectively. Sensitivity analyses suggested that the rate of blindness has the biggest impact on the model results. The interventions may be more cost-effective in populations in which there is a greater risk of blindness. LIMITATIONS The clinical trials did not fully reflect clinical practice. Thirteen additional studies of clinically relevant comparator treatments were identified; however, network meta-analysis was not feasible. The model results are highly uncertain because of the limited evidence base. CONCLUSIONS Two RCTs of systemic adalimumab and one RCT of a unilateral, single dexamethasone implant showed significant benefits over placebo or a sham procedure. The ICERs for adalimumab were estimated to be above generally accepted thresholds for cost-effectiveness. The cost-effectiveness of dexamethasone was estimated to fall below standard thresholds. However, there is substantial uncertainty around the model assumptions. In future work, primary research should compare dexamethasone and adalimumab with current treatments over the long term and in important subgroups and consider how short-term improvements relate to long-term effects on vision. STUDY REGISTRATION This study is registered as PROSPERO CRD42016041799. FUNDING The National Institute for Health Research Health Technology Assessment programme.