A Convenient and Reliable Multi-Class Classification Model based on Explainable Artificial Intelligence for Alzheimer’s Disease
Xiaoqing Xu, Xiang-yuan Yan
Abstract
There is no cure for Alzheimer’s disease (AD) yet. The best treatment available is early detection and early intervention. However, most patients have not been detected in time, let alone timely intervention. So, the World Health Organization (WHO) calls for an increase in the screening rate of AD. Cognitive scores (CS) have great potential in the detection of AD and are easy to obtain, but there are fewer researches based only on CS for classification for screening. Current machine learning models often have good performance but are non-transparent, which makes them difficult to be accepted by physicians. To address the above problems, we propose a solution (RN-SSAS) for data sets with small sample size, multi-category and category-imbalanced, and prove that it is more robust than Cross-Validation (CV), more importantly, it can still achieve classifications under extreme data dilemmas. Then, based on the solution and CS, a multi-class classification model for AD is established, supplemented by single and global instance interpretations based on SHapley Additive exPlanations (SHAP). Finally, our model achieves an F-measure of 0.878 based on four different cognitive scores. In conclusion, our model is convenient, reliable and trustworthy, thus it can contribute to improving the screening rate of AD.