Generalized Additive Neural Networks for mortality prediction using automated and Genetic Algorithms

Abstract

The prediction of mortality has shown to be a challenge for hospital management. To help in this task, metrics were developed to predict the evolution of the disease severity. One of the most commonly used metric in Intensive Care Units (ICUs) is the SAPS II, based on Generalized Linear Models (GLMs). However, the use of the more flexible Generalized Additive Models (GAMs) provide better results when the association between the outcome and the continuous covariates is nonlinear. Neural networks have also been used for prediction namely those based in the Multi Layer Perceptron (MLP) architecture, as, in theory, they are universal approximators to any continuous function. Some studies have shown that their performances are equivalent to GLMs and, naturally, inspired by GAMs, Generalized Additive Neural Networks (GANNs) were proposed. Because the construction of a GANN is based in a subjective decision making process through the analysis of the residuals plots, studies to automate this process emerged originating new methodologies (AutoGANN). However, these are not free from problems when the number of variables is large. Some improvements were then introduced for model selection, such as, a multistep algorithm that allows more than one modification at the same time in GANNs's architecture. Methods described above have correspondence to evolutionary programming as the search of a better result is performed by small modifications, closely resembling the mutation operator. AutoGANN method and Genetic Algorithm were used in order to find optimal models for predicting mortality at an ICU. These models, as well as a MLP model, were compared regarding their predictive and discriminative abilities.