Adaptive splines-based logistic regression with a ReLU neural network

Description

This paper proposes a non-linear binary classification model. Although linear classification methods are very popular in the field of personalized medicine because of their interpretability, they have proven to be too restrictive. Doctors are convinced of the need to quantify threshold effects for better predictions. Nevertheless, non-linear methods that can be found in the state of the art are not able to automate the segmentation of variables (Segmented Logistic Regression) or are difficult to interpret (Random Forests or Neural Networks). We propose a Neural Network that fully realizes a non-linear logistic regression. The score function of the logistic regression, initially linear, is replaced by a piecewise linear function, modeled by spline functions. The particular architecture of this network automates the segmentation of the variables and guarantees its operational relevance as well as the explicability of its calculated predictions.

Abstract

International audience

Additional details