Learning Covariate Relations in Disease Progression Models Using Symbolic Neural Networks

Covariate modeling provides individual predictions of outcomes by disease progression models. Current methodology for mapping covariates onto model parameters is limited by predefined parametric functions which can result in inadequate covariate selection and biased predictions by the final model. Furthermore, present methodology scales poorly to high-dimensional data due to combinatorial limitations. In the present study, a novel method for automation of covariate model identification in disease progression models is described. Symbolic neural networks are used to simultaneously identify the parametric covariate functions and optimize model parameters of a Markov chain. By stepwise pruning of initially fully connected dense symbolic networks, humanly readable functions representing the covariate relations are produced. The presented methodology is applied to a dataset containing disease progression observations for type 2 diabetes patients. Although utilizing fewer covariates, the resulting model demonstrates predictive performance similar to that of a model which was developed on the same data using state-of-the-art modeling methodology.