Search

Effort has been made to find biomarkers for vascular dementia (VaD). Nevertheless, the current findings are typically obtained through statistical tests of group level differences. In clinical practice, however, it is more common to perform individual level inferences, e.g., to determine if a subject is suffering from VaD, which cannot be resolved with statistical analysis. The goal of this study is to develop a method to effectively discriminate early VaD patients from normal controls by combining EEG features with machine learning methods. The EEG signals were recorded from a total of 15 VaD patients and 21 controls during a visual oddball task. Interregional directed connectivity was derived from directed transfer function (DTF) analysis and used as features in classification. Three machine learning methods, including linear discriminant analysis (LDA), error back-propagation (BP) neural network, and support vector machine (SVM) were used as classifiers, and their classification performance was compared. It was found that VaD patients can be effectively identified using the BP and SVM classifiers with high accuracy. In particular, when the SVM classifier was combined with feature selection by Fisher score, it reached an accuracy 86.11%, sensitivity 86.67%, and specificity 85.71%. The area under the curve (AUC, 0.854) indicates a good identification of VaD patients from the normal controls. Since the EEG is noninvasive, inexpensive, and widely available to use, the current study presents a novel clinical application of machine learning methods and could facilitate automatic screening and diagnosis of the VaD at an early stage in future.