Microsimulation modelling of prostate cancer screening in Sweden

Evidence-based health policy may require modelling of different interventions. The choice of model complexity is a trade-off, where simpler models may be easier to describe and calibrate, while complex models may better represent the disease dynamics and lead to more valid predictions. Modelling for health policy is inherently multidisciplinary, with relevant disciplines including epidemiology, clinical medicine, biostatistics, health economics and computer science. As a motivation, we sought to assess the cost-effectiveness of a new prostate cancer screening test – STHLM3. The STHLM3 test uses a combination of biomarkers and self-reported data for prediction. The STHLM3 test can be used as a reflex test after a PSA test. To assess the cost-effectiveness, we needed a model that represents the natural history of prostate cancer. This model was then used to predict the short- and long-term effects of different prostate cancer testing interventions. To achieve this, we developed a framework for event-oriented, discrete event simulation in R and C++ in Study I. The framework included common random numbers, which reduces the Monte Carlo error, and detailed in-simulation reporting for health economic evaluations. In Study II, we extended an older US prostate cancer model to better model for Gleason score. Model inputs included PSA testing, prostate cancer diagnosis, treatment, management and survival. The calibration of the natural history model included both screened and unscreened populations. We initially calibrated the Swedish "Prostata" model using maximum likelihood estimation with non-linear equality constraints. Subsequently in Study IV, we developed a method based on approximate Bayesian computations and Markov chain Monte Carlo methods. The hybrid method provided a more systematic approach to incorporate evidence at different scales while still using known likelihoods. For Study III, we further extended the calibrated model to include costs, health state values and discounting. We calculated the life-time expected costs and effectiveness under different test interventions. We found that the STHLM3 test was cost-effective in Sweden at a reflex PSA threshold of 2 ng/mL. For broader conclusions, first, microsimulation is a challenging computational and scientific task, particularly for calibration and sensitivity analyses. Second, ironically, the depth of the Swedish health and population registers made it easier to invalidate complex models that had been well validated in other populations. The Swedish data can support efforts to improve existing models for cancer screening and, more broadly, other health interventions. Strengths of our approach included: a flexible, lightweight, fast, scalable, open, microsimulation framework for health policy development; calibration of the natural history model to current incidence by Gleason grading and recent survival, whereas most other models have been calibrated to older, PSA-naïve populations; and incorporation of detailed data and estimates, making best use of the available Swedish health and population registers. Limitations of our approach include: imprecise estimates for the effect of prostate cancer testing on mortality; uncertainty in the validity of the natural history model for predictions outside of observed evidence; and uncertainty in the validity of the health state values. For future work, we plan to extend the Prostata model to include magnetic resonance imaging in combination with the newer prostate cancer screening tests