Incorporating human behaviour in simulation models of screening for breast cancer

Abstract

Simulation modelling is widely used in many industries in order to assess and evaluate alternative options and to test strategies or operating rules which are too complex to be modelled analytically. Simulation software has developed its capability in parallel with the growth in computing power since the 1980s. However in practice, the results from the most sophisticated and complex simulation model may not truly reflect what happens in the real world, because such models do not account for human behaviour. For example, in the domain of healthcare simulation is often used to evaluate the outcomes from medical interventions such as new drug treatments. However in reality patients may not complete the course of a prescribed medication, perhaps because they find the side-effects unpleasant. A simulation study designed to evaluate this medication which ignores such behavioural factors may give unreliable results. In this paper we describe a model for screening for breast cancer which includes behavioural factors to model women’s decisions to attend for mammography. The model results indicate that increasing attendance through education or publicity campaigns can be equally as effective as decreasing the intervals between screens. This would have considerable cost implications for healthcare providers.

Introduction

Operational Research (OR) has been applied in the domain of healthcare for more than 40 years. Since the 1960s OR models have been successfully used to assist clinical decision-making, facility location and planning, resource allocation, evaluation of treatments, and organizational redesign (Jun et al., 1999, Fone et al., 2003). Simulation is widely regarded as the technique of choice in healthcare modelling because of its power and flexibility (Davies and Davies, 1994).

We shall focus on one specific application area, the use of simulation models for the evaluation of medical interventions. It is self-evident that individual behaviour can influence health outcomes. For example, adherence to treatment, in terms of taking a drug correctly (or even taking it at all) is a major factor and has been found in clinical studies to be surprisingly low, even amongst people with chronic or life-threatening conditions. A systematic review by McNabb (1997) studied the attitudes of patients with diabetes to all aspects of their treatment, and found that adherence to insulin injections and other medication varied considerably between studies, from as low as 20% in one study to 80% in another. People may choose not to attend screening programmes because they perceive the test as painful, expensive, a waste of time or merely inconvenient. Lifestyle choices such as diet, exercise and smoking can affect health. Therefore any model which ignores these behavioural factors could give unreliable results.

New treatments or interventions are evaluated in terms of their clinical effectiveness, and their cost-effectiveness. Traditionally, clinical effectiveness is assessed using a randomized controlled trial (RCT). In an RCT the test population is divided randomly; some patients receive the new treatment, and others receive either a placebo or the current best available treatment. In the field of medicine the RCT is regarded as the “gold standard”, but it has considerable disadvantages in terms of cost and time. Clearly there will always be a need for clinical trials of new drugs, but simulation modelling can replicate the effects of an intervention in the trial population in a fraction of the time needed for a full-scale RCT, and the model can then be used to conduct experiments which would be unethical or impracticable to carry out in practice, for example restricting the treatment to selected sub-populations, or treating the entire population of a large city. Simulation can provide additional cost-effectiveness measures to aid healthcare planners and managers, who have to choose whether to invest in particular treatments or technologies, some of which are extremely expensive.

In this paper we consider the specific application of simulation modelling for the evaluation of breast cancer screening programmes. Screening refers to the testing of people who are at risk of developing a given disease, with the aim of early diagnosis of that disease. The population screened may be healthy, as in the case we shall consider, the routine screening of middle-aged women for breast cancer. Alternatively, they may already have an underlying condition which puts them at increased risk; for example diabetic patients can be screened for retinopathy, a complication of diabetes which leads to blindness if untreated. For a screening programme to be worthwhile, there must be benefits in early detection, either to the individual themselves in terms of improved prognosis and treatment, or to society in general, for example preventing the spread of infectious diseases.

To be most cost-effective, a screening programme must be designed to target the at-risk population as accurately as possible. Planners must trade off the cost of screening too large a population and/or too frequently, and thus performing many unnecessary tests, against the cost of screening not enough people and/or not often enough, and thereby missing cases. The costs of delivering the screening programme must be weighed against the costs of the disease if undetected. The test itself must trade off cost and pain/inconvenience to the patient against the accuracy of the test.

Simulation models have been used for many years to evaluate the cost-effectiveness of screening programmes. A simulation model has advantages over a conventional RCT, in that typically screening programmes need to be evaluated over the lifetime of the at-risk population and so a trial of 40 or 50 years might theoretically be required. A range of decisions concerning the target population, the screening frequency, the setting (e.g. primary care or hospital), and the test itself can be tested and cost-effectiveness measures derived. These measures can include cost per life year saved, cost per quality-adjusted life year saved (in the case of people with other conditions, or where lack of treatment leads to disability rather than death), cost per case detected, number of screens required to detect one case, and so on. The results should be discounted, since the costs of a screening programme are incurred at the time of screening, whereas the benefits may not accrue until many years in the future.

The paper proceeds as follows. Section 2 presents some statistics on breast cancer incidence and screening programmes, whilst Section 3 provides a literature review of simulation models applied to breast cancer. Section 4 introduces different psychological models used for health behaviours and how they may be incorporated into simulation models is discussed in Section 5. Section 6 describes the simulation model structure built for this study and Section 7 details the model parameters and data sources used. Model validation and experimentation is described in Section 8 and results presented in Section 9 for various scenarios. Section 10 gives results from a sensitivity analysis on the behavioural variables. The paper concludes with a summary of findings in Section 11 and subsequent discussion in Section 12.