Monstrous MCMC: Fully Bayesian Inference in Cyclops for Massive Observational Data Sets

• Activity Number: 15
• Type: Topic Contributed
• Date/Time: Sunday, August 9, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Trevor Shaddox* and Marc A. Suchard
• Companies: UCLA and UCLA
• Keywords: inhomogeneous Poisson process ; GPU ; observational data

Abstract

Emerging national patient claims and electronic health record databases present a new resource for comparing treatment effectiveness and safety. We model adverse drug events as inhomogeneous Poisson processes and rely on the large-scale observational databases maintained by the Innovation in Medical Evidence Development and Surveillance (IMEDS) program to draw inference on these processes. These data encompass millions of lives, thousands of medical products, and hundreds of observations per patient. This allows us to model the inhomogeneity through thousands of health-related covariates. We exploit model averaging to side-step the challenges of working in this dimensionality. We first simulate from the marginal posterior distribution of covariate inclusion using a Laplace approximation. Mode-finding remains the computational bottleneck, which we overcome by finding non-trivial parallelization opportunities and leveraging the many core computational structure of graphics processor units (GPUs). In particular, we employ the GPU to perform fused transformation-reduction operations, producing 30-fold speed-up. These advances make drawing inference feasible at this scale.