Explainable machine learning using SHAP: An application in Alzheimer’s disease diagnosis

Abstract

[en] Machine learning models are a powerful and increasingly ubiquitous tool in modern science, due to their ability to analyse massive amounts of data and extract valuable information. Nonetheless, the increasing complexity of such models usually impedes the interpretation of their decision-making processes. Shapley Additive Explanations (SHAP) is a recently developed explainability method that poses as a suitable candidate to solve this problem. Healthcare is a field where explainability emerges as a critical aspect for the deployment of machine learning techniques. Recently, some studies have started exploring the usage of SHAP on machine learning models for the diagnosis of Alzheimer’s disease (AD). The aim of this work is, firstly, to conduct a theoretical exploration of SHAP to assess its reliability, and then to apply it in a machine learning model for the diagnosis of AD. A theoretical description of SHAP has been done starting from its basis, which relies on results coming from cooperative game theory. A subsequent analysis has been done to study how factors such as feature correlation or feature interaction may affect the method. Numerical tests with synthetic data have been done to illustrate the previous theoretical discussions. Having theoretically presented and discussed the SHAP method, it has been applied in the context of AD diagnosis. Using publicly available data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, a random forest algorithm has been implemented to classify AD patients and healthy controls, using features derived from structural magnetic resonance imaging and amyloid-beta positron emission tomography. This analysis has served as a proof of context of the clinical applicability of SHAP. Further work should be done in order to refine and fully adapt the method to the context of AD diagnosis.