Bayesian Logistic Regression for Medical Claims Data Using CPU and GPU

• Activity Number: 294
• Type: Topic Contributed
• Date/Time: Tuesday, August 2, 2011 : 8:30 AM to 10:20 AM
• Sponsor: IMS
• Author(s): Ivan Zorych*+ and Patrick Ryan and David Madigan
• Companies: Columbia University and Johnson & Johnson Pharmaceutical R&D, LLC and Columbia University
• Keywords: Bayesian regression ; coordinate descent ; logistic ; lasso ; GPU ; CUDA

Abstract

Bayesian logistic regression for medical claims data is a novel statistical approach that possesses the advantages of regression analysis such as being resistant to confounding by co-medication and adjusting for masking effect. Mapping medical claims data into the form appropriate for the regression analysis is an essential step. We consider several ways to represent claims data in the form appropriate for regression. We investigate Bayesian regression models with either Normal or Laplace priors. Analysis of each condition of interest requires fitting a separate regression model. Fitting such a model is a challenging computational task because each dataset contains millions of reports and thousands of covariates. Our numerical approach to logistic regression relies on coordinate descent algorithm. We consider two implementations of this algorithm, traditional central processing unit, CPU, version and a parallel implementation that utilizes graphics processing unit, GPU. The performance of our approach will be illustrated on the simulated and real data.