Feature extraction and classification of EEG signals in autistic children based on singular spectrum analysis

Autism spectrum disorder is a heterogeneous neurodevelopmental disorder involving social, emotional, cognitive, and behavioral disorders. In recent years, with the increasing number of patients with autism, the autistic population has received more and more attention abroad, but the pathogenesis is still unclear. The clinical diagnosis depends on the behavior observation and scale evaluation. Therefore, objective indicators are of great significance for the diagnosis of autism. Electroencephalogram (EEG) is a commonly used neuroimaging technique for its high temporal resolution. This study mainly foucus on the feature extraction of EEG and classification of children with autism based on the analyses of EEG, including the individualized α peak frequency (PAF) and the relative energy of AFi. In this study, we enrolled 30 children with autism (3−6 years old) and 29 age-matched normal children. Resting-state EEG was obtained from each subject with eyes open in a quiet environment. Singular spectrum analysis (SSA) method is proposed to remove artifacts and extract rhythm from EEG signals. The unified α rhythm AFu, was extracted from the EEG signals of normal and autistic children and the classification results were compared by support vector machine (SVM) method. Furthermore, the weighted centroid point was used to explore the individualized α peak frequency (PAF) and rhythmic AFi. The AFi was extracted by SSA method and then repeated and compared. The results showed that the relative energy of AFu was extracted from the EEG data without preprocessing by SSA; the classification accuracy was 81.36%. After preprocessing by SSA, the classification accuracy was increased to 89.83%, which verified the validity of SSA for artifact removing. Using individual the relative energy of α rhythm AFi as the feature, the classification accuracy is reduced to 81.36%, but the classification accuracy rate of 94.92% could be obtained by using individual the relative energy of α rhythm AFi and α peak frequency as common features. The results showed that the abnormal EEG rhythm in autistic children is reflected in two aspects: frequency distribution and power modulation; the α rhythm of autistic children has a low frequency offset and a decrease in relative energy. This study provides a powerful technical means and scientific basis for the auxiliary diagnosis of autistic children from the point of method verification and pathological revealing.