EEG-based K-Nearest-Neighbor And Naive Bayes Classifiers In Online Predictive Tools For Intervention In Mental Illness

The rising prevalence of mental illness due to increasing pressure has negatively impacted people’s health and even their social life functioning. However, deficiencies of existing mental health care in most of the countries including lack of professionals, limited treatment alternatives and inadequate availability hinder prompt alleviation of the situation. Thus there is an urgent call for a system which can predict the mental disorder onset as well as provide timely online intervention. The "Online Predictive Tools for Intervention in Mental Illness" proposed by an international collaborative project of European Union’s Frame Program7perfectly meets the above requests. This paper presents the system architecture, sensor design and application of EEG-based K Nearest Neighbor Classifiers (K=1,2,3) and Naive Bayes Classifier in the project, as well as the calibration trial in China, Swizerland and Spain.In order to predict depressive risk among high stress population based on EEG, the K Nearest Neighbor (K=1,2,3) and Naive Bayes Classifier are adopted in consideration of simple implementation, low computational complexity and good performance on both linear and nonlinear EEG signal features. The performance of EEG signal features, classification accuracy and scalability are evaluated, as well as the performance of Fisher Linear Discriminant Analysis on nonlinear EEG features.Compared to relevant studies, the innovation of the paper lies in:(1) calibration trial in three countries in non-lab environment and depressive risk prediction based on EEG;(2) signifant performance of EEG complexity features in classification;(3) evaluation of calibration accuracy and scalability;(4) relatively satisfactory performance of Fisher Linear Discriminant Analysis on nonlinear EEG features on contrary to relevant studies;(5) selection of EEG signal with stable performance, fitting for similar data set with larger volume.The data analysis results show that EEG signal features can be used to predict depressive risk with successful classification rate over90%for multiple days’average data and over80%for one-day4-second data clips. Besides, the classification successful rates of several EEG signal feature combinations stabalized or rised when data volume increases, maintaining classification successful rates over90%. It suggests that integration of EEG technologies provides a reliable solution for onset prediction of mental illness as well as progress monitoring of mental healthcare.