Comprehensive Risk Prediction Using Interactive Graph-Guided Fussed Lasso Penalty

• Activity Number: 308
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Health Policy Statistics Section
• Author(s): Zhaonan Sun* and Fei Wang and Jianying Hu
• Companies: IBM T.J. Watson Research Center and University of Connecticut and IBM T.J. Watson Research Center
• Keywords: elderly care ; risk prediction ; multi-task learning

Abstract

The ability to identify high-risk and potentially high-needed individuals before they become high cost is crucial to healthcare system. Comprehensive assessment of risks involves simultaneously evaluating and predicting risks from multiple domains. Risks from different domains can be correlated. However, cross-domain correlation is often unclear. We propose a multi-task learning framework to achieve two goals simultaneously: predicting comprehensive risks, and learning the relatedness between risks. In this framework, we propose a interactive graph-guided fussed Lasso penalty, which allows the two components of the model to leverage the strengths from each other. We discuss the connection between the novel penalty and the generalized soft-thresholding. Simulation studies are conducted to demonstrate the advantage of our framework. We apply the proposed framework to the Second Longitudinal Study of Aging. The results provide insights into risk prediction in elderly care. Specifically, we discuss the shared and individual risk factors for medical, functional, cognitive and social domains, and point out possible cross-domain correlation.