Esophageal cancer models overview

This section summarizes similarities and differences among the three CISNET esophageal cancer models: Multi-stage Clonal Expansion model for Esophageal Adenocarcinoma (MSCE-EAC), Esophageal Adenocarcinoma Model (EACMo), and Microsimulation Screening Analysis Esophagus Cancer Model (MISCAN-ESO). These three population-level simulation models include detailed natural history components that have been calibrated and validated using esophageal adenocarcinoma incidence and population data provided by the Surveillance, Epidemiology and End Results (SEER) program. All three are stochastic (random event) models with clinical components that can be used to analyze specific clinical effectiveness for managing patients with Barrett’s Esophagus (BE). Common inputs used in all models include data on SEER incidence, stage-specific survival and mortality, and data on US age-specific incidence and prevalence of BE and symptomatic gastroesophageal reflux disease (GERD) – a known risk factor for BE and EAC. The primary differences among these models lie in the modeling methodologies, which are summarized in Table 1.

Two of the CISNET models (EACMo and MISCAN-ESO) are based on empirical simulations of natural histories, while the third (MSCE-EAC) model uses maximum likelihood methods to calibrate to age-specific EAC incidence data. Calibration is the process by which each model identifies natural history parameters that are most consistent with observed incidence. A likelihood function specifies the probability of the data occurring as a random sample given the model parameters. In the MSCE-EAC model, the likelihood is maximized to optimize the model parameters (initially unknown cell division, death, and mutation rates) that provide the best fit to EAC incidence data.

DesignMulti-stage clonal expansionRecursive health stateRecursive health state
Mathematical model typeBirth-death-mutation processMarkovSemi-Markov
Inclusion of age-period-cohort effectYesYesYes
Explicit modeling of Low Grade Dysplasia (LGD)NoYesYes
Subdivision of BE into long or short segmentsContinuous BE segment lengthYesNo
Considers possible regression of precancerous lesionsYesNoYes
Calibrates to BE targetsYesYesYes
Considers genetic alterationsYesYesNo
Table 1 Similarities and differences among the three EAC models. While there are many similarities, there are important differences, which allow rigorous model-based testing of competing hypotheses. *Recursive states are used to model events that may occur repeatedly over time. Markov models are recursive with transition rates between states that depend only on the current state (but not on prior history).  Semi-Markov models are also recursive, but they generate durations (called sojourn times) in distinct health states instead of modeling transition rates.

The MSCE-EAC model was developed by researchers at Fred Hutchinson Cancer Research Center (FHCRC) in Seattle, Washington. It combines maximum likelihood and multiscale spatial simulation methods to model health states as observations or detection processes built into a detailed tissue- and cell-level model of carcinogenesis.(1-5) The model represents age-dependent development of weekly or more frequent GERD symptoms, with transitions from both GERD and non-GERD pathways to the development of BE. The multiscale spatial methods complement the likelihood-based model through 2-D spatial simulations of BE formation in the esophagus, a two-step cell initiation process, growth of High Grade Dysplasia (HGD) and malignant clones, biopsy sampling, treatment, and clinical diagnosis of EAC.(3-5).

The EACMo model was developed by researchers at Massachusetts General Hospital (MGH) and Columbia University Irving Medical Center (CUIMC). It is a Markov state transition simulation model which simulates a cohort of hypothetical individuals and does not consider the possibility of disease regression. (5-8) Simulated populations enter the model in the Normal Population state and may advance through up to five additional health states, with age, period, and cohort trends applied to transition rates between health states to calibrate to EAC incidence, which increased dramatically during the past 40 years. This model has a microsimulation component that simulates the individual patient's experience during surveillance and treatment.

The MISCAN-ESO model was developed Erasmus University Medical Center and the University of Washington. It is a semi-Markov model which simulates individuals passing through a sequence of health states one at a time and also considers the possibility of disease regression in the health states prior to cancer.  Instead of modeling annual transitions with associated transition probabilities, the MISCAN-ESO model generates durations in states. Two stages of dysplasia are defined — Low Grade (LGD) and High Grade (HGD) — with regression allowed between normal and dysplastic states. From HGD, malignant cells may arise and transform to preclinical localized EAC, which can sequentially progress into Regional and Distant preclinical EAC. (5)

Although the three models differ substantially in their structure, they all generate predictions of incidence and mortality as a function of age and stage of EAC diagnosis. For our recent base case analysis, all three models have been extended to estimate the effectiveness and efficiency of surveillance and treatment for diagnosed BE patients, focusing on the impact of endoscopic ablative treatment (burning away the diseased tissue) on long-term EAC mortality, with risk stratification by gender, age, and dysplastic grade.

The models included in this collaborative program represent diverse modeling perspectives and approaches to synthesizing available evidence to represent the natural history of EAC. These differences will be complementary and will lead to a rich and meaningful comparative modeling experience.


  1. Jeon J, Luebeck EG, Moolgavkar SH. Age effects and temporal trends in adenocarcinoma of the esophagus and gastric cardia (United States). Cancer Causes & Control. 2006;17(7):971-81. [Abstract]
  2. Luebeck EG, Curtius K, Jeon J, Hazelton WD. Impact of tumor progression on cancer incidence curves. Cancer research. 2013;73(3):1086-96. [Abstract]
  3. Curtius K, Hazelton WD, Jeon J, Luebeck EG. A Multiscale Model Evaluates Screening for Neoplasia in Barrett's Esophagus. PLoS computational biology. 2015;11(5):e1004272. [Abstract]
  4. Hazelton WD, Curtius K, Inadomi JM, Vaughan TL, Meza R, Rubenstein JH, et al. The role of gastroesophageal reflux and other factors during progression to esophageal adenocarcinoma. Cancer epidemiology, biomarkers & prevention : a publication of the American Association for Cancer Research, cosponsored by the American Society of Preventive Oncology. 2015;24:1012-23. [Abstract]
  5. Kong CY, Kroep S, Curtius K, Hazelton WD, Jeon J, Meza R, Heberle C, Miller MC, Choi SE, Lansdorp-Vogelaar I, van Ballegooijen M, Feuer EJ, Inadomi JM, Hur C, Luebeck EG.  Exploring the Recent Trend in Esophageal Adenocarcinoma Incidence and Mortality Using Comparative Simulation Modeling, Cancer Epidemiol Biomarkers Prev. 2014 Jun; 23(6): 997-1006. [Abstract]
  6. Hur C, Hayeck TJ, Yeh JM, Richards EB, Spechler SJ, Gazelle GS, et al. Development, Calibration, and Validation of a U.S. White Male Population-Based Simulation Model of Esophageal Adenocarcinoma. PloS one. 2010;5(3):e9483. [Abstract]
  7. Kong CY, Nattinger KJ, Hayeck TJ, Omer ZB, Wang YC, Spechler SJ, et al. The impact of obesity on the rise in esophageal adenocarcinoma incidence: estimates from a disease simulation model. Cancer epidemiology, biomarkers & prevention : a publication of the American Association for Cancer Research, cosponsored by the American Society of Preventive Oncology. 2011;20(11):2450-6. [Abstract]
  8. Hayeck TJ, Kong CY, Spechler SJ, Gazelle GS, Hur C. The prevalence of Barrett's esophagus in the US: estimates from a simulation model confirmed by SEER data. Diseases of the esophagus : official journal of the International Society for Diseases of the Esophagus. 2010;23(6):451-7. [Abstract]