All-Cause Mortality Prediction in T2D Patients

Abstract

Mortality in elderly population having type II diabetes (T2D) can be prevented sometimes through intervention. For that risk assessment can be performed through predictive modeling. This study is part of a collaboration with Maccabi Healthcare Services’ Electronic Health Records (EHR) data, that consists on up to 10 years of 18,000 elderly T2D patients. EHR data is typically heterogeneous and sparse, and for that the use of temporal abstraction and time intervals mining to discover frequent time-interval related patterns (TIRPs) are employed, which then are used as features for a predictive model. However, while the temporal relations between symbolic time intervals in a TIRP are discovered, the temporal relations between TIRPs are not represented. In this paper we introduce a novel TIRPs based patient data representation called Integer-TIRP (iTirp), in which the TIRPs become channels represented by values representing the number of TIRP’s instances that were detected. Then, the iTirps representation is fed into a Deep Learning Architecture, which can learn this kind of sequential relations, using a Recurrent Neural Network (RNN) or a Convolutional Neural Network (CNN). Finally, we introduce a predictive model that consists of a committee, in which two inputs were concatenated, a raw data and iTirps data. Our results indicate that iTirps based models, showed superior performance compared to raw data representation and the committee showed even better results, this by taking advantage of each representations.