Tensor-Factorization-Based Phenotyping using Group Information: Case Study on the Efficacy of Statins

Abstract

To automatically extract medical concepts from raw electronic health records (EHRs), several applications based on machine learning techniques have been proposed. Among the various techniques, tensor factorization methods have attracted considerable attention because tensor representations can capture interactions among high-dimensional EHRs. Most of the existing tensor factorization methods for computational phenotyping are only designed toderive individual phenotypes that approximate the original data. However, deriving grouped phenotypes is desirable because patients form natural groups of interest (i.e., efficacy of treatment and disease categories). In this paper, we propose Supervised Non-negative Tensor Factorization with Multinomial Logistic Regression (SNTFL) to derive grouped phenotypes that are discriminative. We define a discriminative constraint to derive grouped phenotypes and jointly optimize a multinomial logistic regression during the tensor factorization process. Our case study on a hyperlipidemia dataset demonstrates that our proposed method obtains better discrimination on patient groups compared to the baselines and successfully discovers meaningful patient subgroups. © 2017 Association for Computing Machinery.