Online Program

Functional Networks as a Novel Data Mining Scheme for Prediction of Healthcare Outcomes: a Comparative Study with Many Popular Statistical\Data Mining Procedures
Ognian Asparouhov, Chief Scientist of Artificial Intelligence and Data Mining 
*Emad Ahmed El-Sebakhy, Senior Scientist of Data Mining and Artificial Intelligence in Healthcare 
Krassimir Latinski, Scientist Research Associate 
Donghui Wu, MEDai Inc. 

Keywords: Functional networks; data mining; classification; Regression; Neural networks; Healthcare outcomes; Support vector machines

Functional networks scheme has been recently used in modeling both continuous and categorical outcomes. In these networks the functions associated with the neurons are not fixed but are learnt from the available data. This approach has shown its effectiveness in solving wide range of problems, for instance, signal processing, pattern recognition, functions approximations, real-time flood forecasting, bioinformatics and medicine, structure engineering, and other business applications. The purpose of this study is to popularize predictive modeling potential of functional networks to statistical community, especially, for healthcare outcomes. We introduce briefly both classical and some new functional networks learning algorithms and compare the effectiveness’s of functional networks as an alternative approach to well known statistical/data mining procedures for prediction of healthcare outcomes. In this study, we present two groups of experiments (i) prediction of future continuous outcomes - total healthcare cost, total Rx cost, disease specific (asthma, diabetes, mental money for people who had depression,…etc) cost, etc.; (ii) prediction of future categorical outcomes (classification) - disease complications, breast cancer, hospitalization, emergency department visits, etc. In both groups of experiments, we compare the performance of functional networks versus the one of the most common statistical/data mining approaches (regression, support vector machine, K-nearest neighbor, decision trees) using numerous types of data (large/small number of predictors, large/small data set, continuous and/or categorical predictors, etc.