|Category Reference for Presentations|
|AHE||Applied Health Economics||DEC||Decision Psychology and Shared Decision Making|
|HSP||Health Services, and Policy Research||MET||Quantitative Methods and Theoretical Developments|
* Candidate for the Lee B. Lusted Student Prize Competition
Purpose: Accurate biomarkers to routinely assess the effectiveness of treatments for chronic diseases are not always available. Therapies for such conditions are often only effective in a subgroup of patients, where effectiveness is indicated by the rate of disease relapse or death and may or may not be correlated with quality of life on treatment. Medical decision makers must decide what treatments to prescribe for such diseases based on noisy, subjective feedback from patients. We develop a model for choosing between two treatments with stochastic effectiveness: a safe treatment whose effectiveness distribution is known, and a risky treatment which, unknown a priori, could be either superior or inferior to the safe treatment.
Methods: We develop a Bayesian, continuous-time, two-armed bandit model where the alternative treatments give Brownian rewards (health utilities). We model disease relapses according to Poisson processes. If the risky treatment is superior, it leads to relapses at a lower rate than the safe treatment. Unlike classic bandit models which maximize rewards over a deterministic time not influenced by the decisions, our objective is to maximize the patient's cumulative discounted health utilities over the random time interval that ends with the first disease relapse or death, which depends stochastically on the chosen treatment. We apply the model to the problem of choosing between symptom management (safe treatment) or disease-modifying agents (DMA) (risky treatment) for patients with multiple sclerosis (MS), where DMA is considered superior if the patient is a treatment responder and inferior if she is a non-responder.
Results: We find a closed-form analytical solution that provides a threshold probability, representing the current belief of the risky treatment being good, below which it is optimal to choose the safe treatment. For MS, if we assume a constant negative shift in quality of life on DMA due to side effects and maximize cumulative discounted quality-adjusted life years, the optimal threshold for acceptance is 25.8%. If we do not consider DMA side effects and only maximize the discounted time until the next relapse or death, the threshold becomes 6.3%.
Conclusions: By optimally balancing the rewards that the patient receives and the amount of information acquired about the treatment, our model can inform treatment decisions for chronic diseases where patients have unknown responsiveness to treatment.
Purpose: Many compartmental models have projected substantial reductions in HIV prevalence with expanded use of antiretroviral therapy (ART). Often these models simulate HIV disease as progression through 3-5 “health states,” culminating in death; ART slows, but does not reverse, this progression. Our objectives were to develop a dynamic HIV transmission model that could incorporate a more detailed simulation of HIV treatment, and to examine the implications of this framework for long-term epidemic projections.
Result: In the base case, HIV incidence declines rapidly from 2.24/100PY to 1.35/100PY, and prevalence declines from 24% to 22% over 20 years. Eliminating the health gains from ART reduces average life-years with uncontrolled viremia from 7.7 to 6.0y; incidence declines to 1.20/100PY, and 20-year prevalence decreases to 19% (Figure).
Conclusion: Using an HIV microsimulation linked to an SI transmission model, we estimate that immediate ART in South Africa will modestly reduce HIV prevalence. Eliminating the health gains derived from ART – akin to simpler models – likely overestimates this reduction. These results highlight the importance of a detailed accounting of ART's health benefits in models of HIV transmission.
Purpose: To develop and calibrate a cardiovascular disease microsimulation (CVDM) model using the validated cell-based Coronary Heart Disease (CHD) Policy Model and national vital statistics.
Methods: The CVDM was programmed in TreeAge, and states for annual cycles were CVD-free, stroke, coronary heart disease, having both coronary heart disease and stroke, and death. Individual characteristics (sex, age, body mass index, blood pressure, lipid profile, and hypertension treatment, smoking and diabetes) for 35-44 year-olds who were free of CVD were obtained from U.S. National Health and Nutrition Examination Surveys (NHANES) 1999-2008 in order to create nationally-representative simulation cohorts. Annual probabilities of developing the first CVD event or non-CVD death were estimated using multivariate logistic regressions estimated from Framingham Heart Study. Annual risk factor changes and weight gain over time were estimated from NHANES. We calibrated the model by comparing model output with a validated, cell-based Markov model of U.S. adults, the CHD Policy Model, using criteria within 5% of incidence and 1% of mortality rates, as well as visual inspection. Another calibration target was age, sex, and cause-specific mortality rates from 2010 U.S. vital statistics. We simulated 10000 40-year-old male and 10000 female individuals for 30-years or until death. Annual and cumulative CVD incidence, CVD mortality, and non-CVD mortality were computed from the model output.
Results: After CVDM calibration, cumulative 30-year incidence and mortality projections were similar between the CVDM and CHD Policy Model (Table). Within ten-year age intervals between ages 45-74 years, CVDM coronary heart disease and stroke mortality rate estimates were within 1 deaths/1,000 of CHD Policy Model and national vital statistics rates. CVDM coronary heart disease and stroke incidence estimates were within 20 events/1,000 of CHD Policy Model estimates.
Conclusions: Validation is important for positing simulation model accuracy, but nationally representative CVD incidence data are not directly available. An well validated cell-based CVD simulation model can be used in conjunction with national survey and mortality data to validate a new microsimulation model of CVD in U.S. adults.
Methods: Using patient-level data from 7 RCTs representing 4455 patients, we developed a Bayesian hierarchical Weibull regression model to combine data while allowing for trial-specific baseline hazard functions and treatment effects. The final model, derived from backwards elimination, included the main effects of treatment, covariates and the interaction between age and treatment. We performed frequentist evaluation of our prediction model using calibration and discrimination statistics. We performed internal validation using bootstrap samples of the combined data set and external validation using registry data. The model explored patients in 192 subgroups stratified by treatment, age, ejection fraction (EF), New York Heart Association (NYHA) class, QRS, and presence of ischemic disease.
Results: With the borrowing of strength between covariate categories and across trials, our Bayesian hierarchical model allows predictions even for subgroups with small sizes (subgroup sample size ranged from 0 to 200) though with increased uncertainty in such cases. The prediction model had a C-statistic of 0.72 (se=0.01) at year 1 indicating good discrimination and was well calibrated (p=0.99). The C-statistic was slightly smaller at years 2-5 (range: 0.67,0.70), but the model predictions were also calibrated. The same general conclusions were obtained using either internal or external validation data sets. At 5 years, the model predicts the ICD to be more effective in all subgroups. Predicted 5-yr survival with an ICD varied from 29.6% (75+y, NYHAIII, EF<30, QRS>=120, ischemia) to 90% (<65y, NYHAI, EF>=30, QRS>=120, no ischemia), while survival in the control group varied from 20% (75+y, NYHAIII, EF<30, QRS>=120, ischemia) to 84.2% (<65y, NYHAII, EF>=30, QRS<120, no ischemia). The absolute survival benefit ranged from 0.6% to 21.6% across subgroups.
Conclusion: Our findings suggest that over time ICD treatment is more effective in most subgroups relative to non-ICD. Incorporation of this prediction model into a decision-analytic framework will allow exploration of harms/benefits of ICD use in specific subgroups of interest, while also exploring the uncertainty of these findings and the value of additional data acquisition.
Method: Combining the Archimedes model, response surface methodologies, and big data tools, we created a platform that makes evaluating a highly complex disease and healthcare model interactive, while not requiring statistical programming. The Archimedes Model is a large-scale simulation model of chronic disease and healthcare systems. In this method, an individual level dataset is generated with the simulation model and stored in high performance databases for rapid retrieval. In creating the dataset, the simulation model is run for a broad population of individuals, and each individual is simulated repeatedly for a set treatment scenarios specified by a design of experiments. At run time, the user specifies the study subpopulation in terms of clinical parameters. That subpopulation is extracted from the dataset, and response surfaces (e.g. meta-models) are generated. Subsequently, user-specified treatment scenarios can be estimated from the response surfaces, providing estimates equivalent to what the large-scale simulation model would forecast if it were evaluated directly.
Result: We generated an individual level dataset containing 100,000 US Adults with chronic obstructive pulmonary disease (COPD), forecasting biomarkers and outcomes over 5 years, in 50 treatment scenarios, exploring a set of 8 COPD related treatment parameters (such as improved FEV1, FVC, reduced risks of exacerbations, levels of existing treatments, and smoking cessation). The treatment scenarios were specified by a D-optimal design of experiments for continuous and categorical parameters. Using the platform, we can define the study cohort described in the UPLIFT Study of Tiotropium, model the Tiotropium intervention, and obtain estimates in-line with the published trial results in just minutes.
Conclusion: With this platform, large-scale simulation models can be used interactively and without statistical programming. This platform makes generating hypotheses faster, and accelerates the decision making process from months to minutes.