Classification of patients with AD from healthy controls using entropy-based measures of causality brain networks

Machine learning and pattern recognition have been widely used in rs-fMRI data to investigate Alzheimer's disease (AD). However, many previous methods extracted discriminative features based on functional correlations, which may ignore the asynchronous causality influence of neural activities.

We propose a novel method for AD diagnosis using Sample Entropy to measure the neural complexity of the brain causality network. Granger Causality analysis with a sliding time window was applied on rs-fMRI data of 29 AD patients and 30 cognitive normal (CN) controls to compute the whole brain's causality series. We further grouped these causality series into clusters by agglomerative hierarchical clustering algorithm and computed Sample Entropy of the clusters as the classification features.

We explored four different classifiers, i.e., XGBoost, SVM cluster, Random Forest, and SVM, based on the above features. An accuracy of 89.83%, with a sensitivity of 90.00% and a specificity of 89.66%, was achieved with the optimal feature subsets using the SVM classifier.

With the same dataset, the performances of the proposed method were generally higher than those of conventional methods for AD classification based on Pearson's correlation network, dynamic Pearson's correlation network, High-order correlation network, and causality correlation network.

Our method demonstrates the measure of Sample Entropy with causality connection as a powerful tool to classify AD patients from CN controls, and provides a deep insight into the neuropathogenesis of AD.