Classification of acute ischemic stroke EEG signal using entropy-based features, wavelet decomposition, and machine learning algorithms

Stroke is one of the most leading causes of death and disability in the world as well as in Indonesia. Almost 85% of stroke patients suffer from Acute Ischemic Stroke (AIS). They need to be early diagnosed to improve stroke treatment. The most common tools that have been widely used in diagnosing strokes are Computed Tomography Scans (CT Scans) and Magnetic Resonance Imaging (MRI). Electroencephalography (EEG) analysis has been widely studied in identifying stroke disease due to its relatively low cost and non-invasive characteristics. The study is aimed to classify the severity of AIS patients using EEG signals into four classes: normal, minor, moderate, and severe. This study was conducted in Rumah Sakit Pusat Otak Nasional (RSPON, National Brain Center Hospital), Jakarta, and acquired 32-channel EEG data recordings, CT-Scan images, and NIHSS scores. The total subject participated in this study was 57 subjects: 35 male subjects and 22 female subjects with age ranges 40 – 60 years old. Shannon Entropy (SE), and Log Energy feature-based EEG from alpha, beta, theta, delta, and gamma sub-bands were extracted and evaluated as the EEG signals are complex, non-stationary, and non-linear. These features were combined with Delta to Alpha Ratio (DAR) and Delta Theta to Alpha Beta Ratio (DTABR) that were extracted using wavelet decomposition. All of these features were proceeded using three different classifiers: k-Nearest Neighbors, Decision Tree, and Naïve Bayes classifier to compare their performances. Besides classifiers, we also used three different sets of features: All features; Shannon Entropy; Log Energy; Shannon Entropy, and Log Energy features as training inputs. The highest accuracy was yielded by Decision Tree using Shannon Entropy feature, which yields 83% accuracy. This system would be expected to be used widely in type-C hospitals in Indonesia.