SLIDE 2 NOTES: Elements of the Miscan Model’s structure are shown in this frame. The micro-simulation model treats time as a continuous variable, with discrete events occurring along the time line. So, for example, the appearance of a polyp of a particular size is a discrete event. (The model identifies three sizes.) But the dwelling time of such a polyp before it progresses to a larger size can be any length. Probability functions determine when the transition from one state to another will occur.
By parallel universe, we mean that the model first generates a complete simulated, or hypothetical, population of individuals with their complete natural history in the absence of screening. The results of that baseline simulation, with all pertinent clinical details, are stored. Then, the model begins again with the same population as before, but this time a particular screening strategy is imposed. Assumptions about the effect of screening on the clinical course of each patient determine a new set of simulated results. The net effect of the screening strategy is determined per life history by the changes that occur compared with the baseline.
The model provides outputs on a real population in any specific calendar year. Most of our published work has taken that approach. However, the model is flexible in that a specific cohort of individuals (such as 50 year old men) can be followed throughout the rest of their lives. For the pre-workshop exercise, we did adapt the model to a cohort structure.
SLIDE 3 NOTES: The model is also a multi-disease model. By that I mean two things. First, although the model is intended to simulate a general population, it recognizes the inherent variation among individuals in the risk of developing adenomas and cancer. Thus, each individual has his own probability of acquiring one more adenomas.
Second, adenomas are simulated as distinct occurrences in the same person. One or more can occur simultaneously or over time. For example, a person’s simulated natural history might be programmed to develop a non-progressive adenoma in the rectum at one time, and another progressive adenoma in the transverse colon at another time. Each of these events will have their own independent history until a first cancer is detected in that simulated individual.
SLIDE NOTES 4: This chart shows the disease states and transitions that are included in the model. For example, a person with no adenomatous lesions may develop a small adenoma, which may or may not progress into cancer.
The length of time required for each person to travel through each phase in the adenoma/carcinoma sequence is determined by the parameters of a probability distribution for each disease stage and for each transition that is specified as part of the model.
The model as currently quantified does not permit a cancer to derive directly from a small adenoma. The adenoma must first progress to a medium or large size. Once the transition to cancer occurs, the lesion progresses through four stages. In any stage there is a chance, based on an assumed probability of detection, that the lesion will be diagnosed and become a clinical case. Once clinical, the lesion will or will not cause the patient’s death from colorectal cancer with a specific probability specific to each stage.
SLIDE 5 NOTES: The parameters used to generate each hypothetical person’s history were derived from a variety of sources, including data produced in clinical and epidemiological studies and, when no reliable data were available, expert opinion.
We started with data on the US population by age and calendar year. We also used SEER CRC incidence data from the late 1970’s before screening was prevalent in the USA. We also used the extant autopsy studies to estimate adenoma prevalence, including the prevalence of multiple adenomas in the same individual.
SLIDE 6 NOTES: This chart provides an overview of which parameters were based on direct estimates and which resulted from a statistical fit procedure. Those based on assumptions are the ones that we must investigate when we validate the model, because they are the ones for which there are inadequate data from screening studies at present. They include the duration and distribution of the preclinical state and, closely connected, the transition probabilities from one preclinical state to another.
We assumed that it takes, on average, 20 years from the start of a small adenoma to a clinical cancer. We have attempted to validate this assumption with adenoma endoscopy studies, but more such studies are needed for us to be fully confident in that assumption. Another assumption regards the correlation among durations in different preclinical disease states. For now, we assume that if an individual progresses rapidly through the small adenoma state, he or she has a high probability of progressing rapidly through later states. We also assume a positive correlation among the durations of preclinical cancerous states.
At present, the results of screening tests are assumed to be independent of earlier or later tests, but we do not have data to support or reject this assumption. We also assumed that the same stage-specific survival holds for cancers detected through a screening procedure as for those detected through presentation of clinical symptoms. Thus, all the gain from early detection of cancer through screening results from a shift in the stages at which cancers are detected. Better data are needed to determine whether this assumption is justified.
SLIDE 7 NOTES: To arrive at realistic assumptions about the degree of compliance that can be expected from a particular screening strategy, we relied on data reported from clinical trials and population surveys. Population surveys are problematic as a source of compliance estimates, because the population has a choice of 3 or more different screening technologies for colorectal cancer screening, and the surveys do not ALWAYS provide detailed information on the specific test received. Applying compliance rates found in surveys to a screening strategy involving only one procedure (e.g., sigmoidoscopy every 5 years) could result in errors.
Our model assumes that the probability of compliance with the screening strategy at any screening round depends on the person’s compliance history in the previous screening round. Compliance in one round implies, in the model, a higher probability of compliance in the successive round. The probability of attending screening when attended previously is four times the probability when not attended the previous round. The total percentage of people who are attenders is constant over time.
Compliance with all diagnostic follow-up and post-adenoma detection surveillance is assumed to be 100 percent.
SLIDE 8 NOTES: Our assumptions about the costs of screening, follow-up and surveillance procedures are shown elsewhere in the workshop summary. The costs of CRC treatment vary with the number of months since the detection of a CRC. Initial treatment costs cover the first six months; after that, continued CRC-related costs are charged over the remainder of the life, until the last six months, when costs of terminal care are included.
However, for the pre-workshop exercise, because individuals would be followed only until age 85, we did not include terminal care costs, even when a simulated person died before the age of 85.
We include only those costs associated with treating CRC. Unrelated medical care costs are not included, nor are any non-medical costs, such as transportation or work losses.
SLIDE 9 NOTES: We have attempted to validate the model’s predictions by comparing them with the results of several screening and surveillance trials. In comparison to the National Polyp Study (a trial of post-polypectomy surveillance) Miscan predicted more cancers in the years following polypectomy than was reported in the trial (Loeve et al., 2004). In comparison to the Kaiser Flexible sigmoidoscopy study (Levin and Palitz, 2002) we also found too many cancers during the 5 years following a negative sigmoidoscopy (van Ballegooijen et al., 2002). So, we predicted too many cancers during follow up in individuals both with and without adenomas, while our total risk for cancer was calibrated to the total population. To explain the sigmoidoscopy data, we need to adjust our model so that risk is shifted away from individuals with few or now adenomas. Consequently, individuals with adenomas (as in the National Polyp Study) on average will have an increased risk. But this is not compatible with the National Polyp Study, where we already simulated too many cancers. On the other hand, since neither of these two studies was randomized before entry, it cannot be ruled out that the study groups consist of relatively low-risk individuals through self-selection (in the sigmoidoscopy study) or through clinical selection (in the NPS). This shows why we badly need randomized studies for further validation.
We have been able to reproduce the results of the Minnesota trial. Our model predicted a 34.6 percent reduction in morality from annual screening and a 20 percent reduction from biennial screening. In the Minnesota trial the observed reductions were 33 percent and 21 percent, respectively. Although we reproduce the mortality reduction found in that trial, the model did not predict the stage distribution in follow-up rounds so
well. The model predicted a much more favorable stage distribution than was found in the Minnesota trial. That failure may suggest problems if we are to use the model for improved tests or other strategies. We are therefore currently attempting to validate the model against the three big FOBT trials together, combining the data that are available from each one.
SLIDE 10 NOTES: No notes.
SLIDE 11 NOTES: No notes.
REFERENCES
Levin TR, Palitz AM. 2002. Flexible sigmoidoscopy: An important screening option for average-risk individuals. Gastrointest Endosc Clin N Am. 12(1):23–40.
Loeve F, Boer R, Zauber AG, Van Ballegooijen M, Van Oortmarssen GJ, Winawer SJ, Habbema JDF. 2004. National Polyp Study data: Evidence for regression of adenomas. Int J Cancer. 111(4):633–639.
van Ballegooijen M, et al. 2002. Validation Study of the Miscan-Colon Program Sith Data From the CoCaP Sigmoidoscopy Program. Final Report to NCI. Rotterdam, NE: Erasmus Medical Center.