Robust stress classifier using adaptive neuro-fuzzy classifier-linguistic hedges

Recent studies show that chronic stress exposure can induce a long list of diseases that are prevalent in human body. In this paper, researchers work on measuring and analyzing stress level using human biosignal, electrocardiogram (ECG). First, a few preprocessing steps and different analysis domains is done onto the raw data signals to clean and extract any and every relevant features found in ECG signal. A Linguistic Hedges concept on fuzzy feature selection method is then proposed to select unique patterns from the listed heart rate variability features. From the extracted list of features, a neuro-fuzzy classifier (ANFC-LH) is used to classify the data points into 2 classes, high arousal and low arousal, high arousal indicating stress feature. Then a comparative study using different classification methods, including Multilayer Perceptron, k-Nearest Neighbor, and Linear Discriminant Analysis are used to determine the most relevant feature specifying high stress level. Comparing to MLP, kNN, and LDA, ANFC-LH achieved the highest recognition rate. This research paper also shows the effects of using dimension reduction methods on classification algorithms where the result of kNN and LDA improved about 20% when applied with dimension reduction method, however, MLP recognition rate deteriorates about 50% when classifying data point after dimension reduction.