Multilevel Assessment of Mental Stress using SVM with ECOC: An EEG Approach

Abstract

Mental stress has been identified as one of the major contributing factors that leads to various diseases such as heart attack, depression and stroke. To avoid this, stress quantification is important for clinical intervention and disease prevention. This study aims to investigate the feasibility of exploiting Electroencephalography (EEG) signals to discriminate between different stress levels. We propose a new assessment protocol whereby the stress level is represented by the complexity of mental arithmetic (MA) task for example, at three levels of difficulty, and the stressors are time pressure and negative feedback. Using 18-male subjects, the experimental results showed that there were significant differences in EEG response between the control and stress conditions at different levels of MA task with p-values < 0.001. Furthermore, we found a significant reduction in alpha rhythm power from one stress level to another level, p-values <0.05. In comparison, results from self-reporting questionnaire NASA-TLX approach showed no significant differences between stress levels. In addition, we developed a discriminant analysis method based on multiclass support vector machine (SVM) with error-correcting output code (ECOC). Different stress levels were detected with an average classification accuracy of 94.79%. The Lateral Index (LI) results further showed dominant right prefrontal cortex (PFC) to mental stress (reduced alpha rhythm). The study demonstrated the feasibility of using EEG in classifying multilevel mental stress, and reported alpha rhythm power at right prefrontal cortex as a suitable index.