From Risk Prediction to Risk Factors Interpretation. Comparison of Neural Networks and Classical Statistics for Dementia Prediction

TL;DR

This paper compares neural networks and classical statistical methods for predicting dementia, focusing on interpretability of risk factors, with neural networks excelling in high-dimensional data and classical methods providing clearer factor roles.

Contribution

The study provides a comparative analysis of neural networks and classical statistics in dementia prediction, emphasizing interpretability and performance differences.

Findings

Neural networks perform better with high-dimensional predictors.

Classical statistics offer clearer interpretation of risk factors.

Both methods are effective for disease onset prediction.

Abstract

It is proposed to investigate the onset of a disease D, based on several risk factors., with a specific interest in Alzheimer occurrence. For that purpose, two classes of techniques are available, whose properties are quite different in terms of interpretation, which is the focus of this paper: classical statistics based on probabilistic models and artificial intelligence (mainly neural networks) based on optimization algorithms. Both methods are good at prediction, with a preference for neural networks when the dimension of the potential predictors is high. But the advantage of the classical statistics is cognitive : the role of each factor is generally summarized in the value of a coefficient which is highly positive for a harmful factor, close to 0 for an irrelevant one, and highly negative for a beneficial one.