Automated Detection of Alcohol Related Changes in Electroencephalograph Signals

It is difficult to detect and treat alcoholism, because statistics show that statements from patients about their drinking habits are unreliable and diagnosable symptoms appear only in advanced stages of the disease. To address this problem, we propose an automatic system that characterizes alcohol related abnormalities in Electroncephalography (EEG) signals. This system enables clinicians, patients and all other people involved to manage the condition better. Furthermore, it provides deeper insights into the phenomena and thereby it reveals important clinical information about alcohol related changes in EEG signals. For this work, we adopt the widely held, and evidence supported, belief that EEG recordings are fundamentally nonlinear. As direct consequence, the nonlinear feature of Higher Order Spectra (HOS) cumulants was used to extract information about alcohol related changes from the EEG signals. The decision whether or not a particular EEG signal shows alcohol related changes, was established with six different classification algorithms: Decision Tree (DT), Fuzzy Sugeno Classifier (FSC), K-Nearest Neighbor (KNN), Gaussian Mixture Model (GMM), Naive Bayes Classifier (NBC) and Probabilistic Neural Network (PNN). To establish the functionality, we tested the proposed diagnosis support system with 300 EEG data sets. The individual classification algorithms achieved different accuracy values, they ranged from 77% (NBC) to 92.4% (FSC). The (FSC) classification result supports our thesis that HOS based cumulants features can be used to discriminate alcohol and normal EEG signals. The fact that there was a wide range of classification accuracies supports our decision to test four different classification algorithms.