סמינר מחלקתי - ליאור תורג'מן

Abstract:

In this talk, I will present a novel approach for predictive modeling, using a mixed-ensemble classifier. The approach integrates a C5.0 tree as the base ensemble classifier, and a support vector machine (SVM) as a secondary classifier. The tree base classifier enables exploratory knowledge discovery about the learned database.  The flexibility of the secondary SVM algorithm, in terms of controlling the width of the margins around the separating hyperplane, and in terms of classification error, allows for a controlled-sensitive classification of the minority class. By implementing our method for predicting hospital readmission of Congestive Heart Failure (CHF) patients in the Veterans Health Administration (VHA) hospital system of Pittsburgh, we were able to overcome some of the limitations of both C5.0 and SVM, as well as to increase the classification accuracy for the minority class, particularly when strong predictors are not available. I will also present a new approach of identification of readmission risk factors during the hospitalization process, by fitting a Coxian phase-type distribution to their length of stay (LOS) data, and analyzing the strength of the connections among patient characteristics within each extracted  state in the latent Markov process. Some possible research directions of applying the suggested methods for other fields of application-driven research will be also presented.

Bio :

Lior Turgeman is a postdoctoral researcher at the Joseph M. Katz Business School, University of Pittsburgh, where he has been since 2014. Lior completed his Ph.D in Engineering at Bar-Ilan University on 2013. His research interests lie in the area of stochastic modeling, systems engineering, data mining and predictive analytics, ranging from theory to design to implementation, with a focus on application-driven systems. In recent years, he has focused on developing better techniques of data analysis, and optimization, of healthcare and biomedical measurement systems, particularly in case where only partial knowledge of their state is given. He has collaborated actively with researchers in several other disciplines of medicine, industrial engineering, and computer science, as well as with the Pittsburgh Veterans Engineering Resource Center (VERC).