51 ITERATIVE CALIBRATION OF MONTE-CARLO MARKOV MODELS USED FOR EVALUATING THE IMPACT OF UPDATING TRADITIONAL CARDIOVASCULAR RISK PREDICTIONS WITH NOVEL RISK MARKERS

Thursday, October 18, 2012
The Atrium (Hyatt Regency)
Poster Board # 51
INFORMS (INF), Quantitative Methods and Theoretical Developments (MET)
Candidate for the Lee B. Lusted Student Prize Competition

Bob JH Kempen, MSc1, Bart S. Ferket, MD1, Ewout W. Steyerberg, PhD2, Oscar H. Franco, MD, PhD, FESC, MFPH3, Wendy Max, PhD4, Kirsten E. Fleischmann, MD, MPH5 and M.G. Myriam Hunink, MD, PhD6, (1)Erasmus MC, Rotterdam, Netherlands, (2)Department of Public Health, AE 236, Rotterdam, Netherlands, (3)Erasmus MC, University Medical Center Rotterdam, Rotterdam, Netherlands, (4)University of California, San Francisco, San Francisco, CA, (5)UCSF Medical Center, San Francisco, CA, (6)Erasmus University Medical Center, Rotterdam, Netherlands

Purpose:    To present and validate an iterative method to calibrate Monte-Carlo Markov models used for evaluating the impact of updating traditional cardiovascular disease (CVD) risk predictions with novel risk markers.

Method:    We developed a Monte Carlo-Markov model with three health states: 1) alive and CVD-free, 2) post-CVD, and 3) death. One-year transition probabilities were based on the Framingham 30-year cardiovascular risk function, which include traditional factors and takes into account competing non-cardiovascular death. We updated the Framingham risk function with the CT coronary calcium score, ankle-brachial index, high-sensitivity C-reactive protein and carotid intima-media thickness, by extending the original linear predictor with the respective adjusted beta-coefficients from meta-analyses. Individual risk profiles, containing information on the traditional and 4 novel risk factors were taken from 3,736 asymptomatic subjects of the National Health and Nutrition Examination Survey (NHANES).    We assumed that the average CVD risk based on the traditional risk factors alone would not change as a result of the addition of the novel risk factors. Using a cycle length of 1 year, we calculated the uncalibrated 1-yr CVD risk by using the hazard of year 1 and the extended linear predictor. We then added a fixed term to the extended linear predictor such that the average 1-yr CVD risk of all 3,736 individuals equalled the 1-yr CVD risk based on the original Framingham risk function. Using the calibrated first year probability of CVD, we simulated the 3,736 individuals and tracked which individuals experienced a CVD event or competing death. The remainder of these individuals were used to recalibrate the 1-yr CVD risk of year 2. This was repeated for 30 years.    Finally, we compared the 30-year CVD risk simulated by the calibrated Monte-Carlo Markov model for the 3,736 individuals with the predicted 30-year CVD risk based on the original Framingham risk function.  

Result:    The average CVD risk at year 1-to-30 of the 3,736 subjects (median age 53 years ICR 46 - 63, 48% male) simulated by the calibrated Monte-Carlo Markov model matched the predicted CVD risks based on the original Framingham risk function (figure 1).

Conclusion:    We presented a valid method to calibrate Monte-Carlo Markov models which can be used to evaluate the impact of updating traditional CVD risk functions with novel risk markers. figure tech abstract