Explainable Decision Tree-Based Screening of Cognitive Impairment Leveraging Minimal Neuropsychological Tests

Abstract

Cognitive impairment detection is challenging because it primarily depends on advanced neuroimaging tests, which are not readily available in small cities and villages in India and many developing countries. Artificial Intelligence (AI)-based systems can be used to assist clinical decision-making, but it requires advanced tests and a large amount of data to achieve reasonable accuracy. In this work, we have developed an explainable decision-tree-based detection model, which serves as a powerful first-level screening on a small basic set of cognitive tests. This minimum set of features is obtained through an ablation study. The Alzheimer’s Disease Neuroimaging Initiative (ADNI) archive provided the data for this study. We obtained 93.10% accuracy for three classes: Cognitive Normal (CN), Mild Cognitive Impairment (MCI), Alzheimer’s Disease (AD), and 78.74% accuracy for five classes: CN, Significant Memory Complaints (SMC), Early-MCI (EMCI), Late-MCI (LMCI), AD using Extreme Gradient Boosting (XGBoost) which is comparable to the accuracy of state-of-the-art methods which use a more sophisticated and expensive test like imaging. Current research pays little attention to explainability and is primarily concerned with enhancing the performance of deep learning and machine learning models. Consequently, clinicians find it challenging to interpret these intricate models. With the use of Tree Shapley Additive Explanations (TreeSHAP) values and Local Interpretable Model-agnostic Explanations (LIME), this work intends to give both global and local interpretability respectively. Moreover, we highlight the use of top-2 metrics (Accuracy, Precision, and Recall), which significantly improves corner cases and helps the clinician streamline diagnosis.