To Register      SMDM Homepage

Tuesday, 19 October 2004 - 11:15 AM

This presentation is part of: Oral Concurrent Session A - Simulation

DEVELOPMENT OF A NATURAL HISTORY MODEL OF LUNG CANCER

Pamela M. McMahon, BS1, Karen M. Kuntz, ScD2, Milton C. Weinstein, PhD2, Jane C. Weeks, MD3, Alan M. Zaslavsky, PhD4, and G. Scott Gazelle, MD, MPH, PhD5. (1) Massachusetts General Hospital, Harvard University, Institute for Technology Assessment, Boston, MA, (2) Harvard School of Public Health, Harvard Center for Risk Analysis, Boston, MA, (3) Harvard School of Public Health, Health Policy and Management, Boston, MA, (4) Harvard Medical School, Health Care Policy, Boston, MA, (5) Massachusetts General Hospital, MGH-Institute for Technology Assessment, Boston, MA

Purpose: To develop a natural history model of lung cancer to underlie a Lung Cancer Policy Model (LCPM). Unlike published models of lung cancer screening, the LCPM avoids reliance on observed stage shifts as the mechanism for survival gains, thereby allowing estimation of screening effectiveness in populations not represented in trials.

Methods: The natural history model is a state-transition Monte Carlo model. Transitions among the states (general population free from diagnoses, clinical staging, follow-up of indeterminate lesions, and dead) are functions of underlying cancerous or benign lesions, detection via symptoms or incidental imaging exams, test characteristics, and risk factors. Input parameters are based on national survey and vital statistics data, literature estimates of lung cancer doubling times, and incidence of benign lesions observed in CT screening trials. Tumor growth is modeled with Gompertz functions, and metastasis depends on tumor characteristics and individual variation. Parameters for unobservable events (e.g., metastasis) are derived by calibration to multiple endpoints from tumor registries, trials, and the literature. For each simulated individual, characteristics include: demographics; smoking history (cigarettes/day, years of smoking, years since quitting, secondhand smoke exposure); and disease characteristics (locations, sizes, doubling times of ≤3 lung cancers and ≤3 benign lesions, metastases, and nodal involvement). Effects of demographics and smoking on competing risks were estimated from a Bayesian evidence synthesis of individual survey and population data with published cohort studies: coefficients incorporate 1st and 2nd order uncertainty and preserve correlations between predictors.

Results: The first phase model (onset of cancer only) produces results consistent with SEER estimates (incidence by stage, histological type, and size), autopsy studies, screening trials, clinical experience, and cohort studies (e.g., lung cancer in lifelong non-smokers). Age-specific incidence estimates respond in predictable ways to changes in inputs (e.g., higher smoking rates lead to more lung cancers). Calibration to SEER incidence estimates suggests there is a large reservoir of undetected lung cancers.

Conclusions: Our model integrates multiple data sources and is calibrated to tumor registry data, and provides insights into the natural history of lung cancer. Unlike published models, our model design will yield estimates of lead, length, and overdiagnosis biases.


See more of Oral Concurrent Session A - Simulation
See more of The 26th Annual Meeting of the Society for Medical Decision Making (October 17-20, 2004)