Performance Comparison of Machine Learning Algorithms in Dementia Classification Using Electroencephalogram Decomposition

Dementia has significant social and economic impacts on direct medical, social, and informal care costs. An effective diagnostic method may provide an early diagnosis of the disease, giving a leeway to tackle the burden. This paper presents a performance comparison of five machine learning techniques for dementia classification based on a feature extraction method using cardinal-spline empirical mode decomposition (CS-EMD) of electroencephalogram (EEG). The comparison was validated using repeated k-fold cross-validation and the Wilcoxon signed-rank test. The top two classifiers are artificial neural network (ANN) and support vector machine (SVM), with no statistically significant difference in performance in terms of overall accuracy. The ANN and SVM achieved average classification accuracies of 93% and 88.42%, respectively, outperforming discriminant analysis, the K-nearest neighbors, and the naïve Bayes classifier in the repeated k-fold CV. Furthermore, the ANN classifier in a train/test split test produced higher accuracy and sensitivity rates, but lower specificity and precision than the SVM. The results show good potential for the CS-EMD-based features of EEG decomposition and machine learning classifiers, such as ANN and SVM, in dementia diagnostic applications.