Design of a Biosignal Based Stress Detection System Using Machine Learning Techniques

This study represents a design of a detection system of stress through machine learning using some available bio signals in human body. Stress can be commonly defined as the disturbance in psychological equilibrium. Stress detection is one of the major research areas in biomedical engineering as proper detection of stress can conveniently prevent many psychological and physiological problems like cardiac rhythm abnormalities or arrhythmia. There are several bio-signals available (i.e. ECG, EMG, Respiration, GSR etc.) which are helpful in detecting stress levels as these signals shows characteristic changes with stress induction. In this paper, ECG was selected as the primary candidate because of the easily available recording (i.e. several mobile clinical grade recorders are available now in the market) and ECG feature extraction techniques. Another advantage of ECG is that respiratory signal information can also be detected form ECG which is known as EDR (ECG derived Respiration) without having separate sensor system for respiration measurement. Features of ECG signals are distinctive and collection of the signals is cost-efficient. From ECG we derived RR interval, QT interval, and EDR features for the development of the model. For the implementation of a supervised machine learning (SVM) method in MATLAB, Physionet's "drivedb" database was used as the training dataset and validation. SVM was chosen for classification, as there are two classes of labeled data; `stressed' or `non-stressed'. Several SVM model types were verified by changing the feature number and Kernel type. Our results showed an accuracy level of 98.6% with Gaussian Kernel function and using all available features (RR, QT and EDR), which also emphasizes the importance of respiratory information in stress detection through Machine Learning.