Investigating Effects of the Window Size on Stress Detection Through Machine Learning

Recently, stress detection has become an important research area with the increased usage of sensor technologies. In this study, the effect of segmenting the dataset with various window sizes on stress classification was studied. Publicly-accessible dataset, Wearable Stress and Affect Detection (WESAD) was employed and stress classification was performed. Only the data of the chest sensors namely electrocardiogram (ECG), electrodermal activity (EDA), electromyography (EMG) and temperature (TEMP) were used. Sensor data were segmented using the sliding window technique with time window sizes ranging from 10 to 120 seconds. Then, the segmented data were used in the extraction of temporal and spectral features. The derived temporal and spectral features were employed to train and test the classification algorithms. In this study binary classification method was used. Decision Tree (DT), k Nearest Neighbors (k-NN) and Linear Support Vector Classification (LinearSVC) were used as classification algorithms. The data tested with classification algorithms were evaluated according to metrics namely accuracy, F1 score, precision and recall. The results provide guidance for evaluating the implications of various window sizes on classification efficiency and also provide important findings for determining the most appropriate window size parameters and machine learning algorithms for stress classification.