Quantitative Modeling Software with Applications to Medical Decision Making - Project Summary/Abstract
In recent years, health care systems and health care providers have made concerted efforts to practice
evidence based medicine to provide patients with the best available information when making choices about
their medical decisions. However, these decisions are often complex with many uncertainties and potential
outcomes – some beneficial others dire. A popular tool used to help identify best treatment strategies is a
decision tree, which outlines a patient's potential outcomes given an initial medical choice. Complex decision
trees are evaluated via Monte Carlo microsimulation to trace a patient's path through the tree. This movement
is inherently stochastic because outcomes are probabilistic; however, when the microsimulation is repeated
many times, it provides the probability of each associated outcome resulting from the initial medical decision.
From this probability distribution, quantitative measures associated with each medical decision can be
calculated including beneficial as well as adverse events, life years, quality-adjusted life years (a generic
measure of disease burden), and others. When outcome costs are known and incorporated into the model,
cost-effectiveness analysis (CEA) can be used to readily compute the relative costs, effectiveness, and
incremental cost-effectiveness for each health outcome. We propose to add functionality to the mathematical
modeling software Berkeley Madonna to allow users to build decision trees and carry out Monte Carlo
microsimulations on these trees (Aim 1). Berkeley Madonna's interface was designed to gently introduce
students from non-technical fields into mathematical modeling by using a simple syntax and graphical images
to construct sophisticated equations. We will leverage this easy-to-use interface to introduce medical
researchers to microsimulation. The software will be adapted to build decision trees with built-in functions and
customized graphics specific to this field, including measures from CEA. Patients moving through a decision
tree using microsimulation must be simulated hundreds of thousands to millions of times to arrive at statistically
significant outcome probabilities, and these simulations are computationally intense often requiring months of
computer time. We will harness the power of graphics processing units (GPUs) to parallelize these simulations
to achieve tremendous speedups compared to commercially available software, which have not taken
advantage of these hardware capabilities. We show that a simple Monte Carlo microsimulation can be
simulated 700x faster on a GPU compared to a CPU, and we will optimize the code to fully realize these
speedups when simulating complex decision trees (Aim 2). Successful completion of our goals will provide a
powerful research tool to the medical decision making field, which will positively impact health outcomes
research. We expect that our software will be particularly beneficial to the cardiovascular research community,
which has a history of practicing evidence-based medicine that includes simulation modeling.
