Generalized additive neural network with flexible parametric link function: model estimation using simulated and real clinical data

摘要

Artificial neural networks have been proposed in medical research as an alternative to some regression models such as the generalized linear models, being the multilayer perceptron (MLP) the most used architecture. However, inspired in the generalized additive models (GAM), recent studies proposed the more transparent generalized additive neural network (GANN) architecture. In fact, while a MLP may be seen as a black box in which the effect of a variable on the outcome is not clear, a GANN has the advantage of being able to study objectively, through a graphical approach, the effect of an input variable on a certain outcome of interest [9]. In this study, the GANN's architecture was updated, considering some features already available in the GAM, namely the use of a flexible parametric link function based on the Aranda-Ordaz transformations family for a binary response. Also, the interpretability was improved by obtaining the partial functions with the corresponding confidence intervals through the bootstrap method. The performance of the proposed model was evaluated with simulated data and further applied to a real clinical dataset.