Metody extrakce příznaků a umělé inteligence pro klasifikaci a kvantifikaci stresu v rámci Esport

Abstract

The diploma thesis deals with the classification of stress and non-stress conditions during Esport activities based on biological signals of electrodermal activity (GSR/EDA), heart rate (HR) and RR intervals. Using the open-source library „tsfresh“, features were extracted from the time series of signals, which were then reduced using statistical significance tests. For the classification of stress and non-stress conditions, three classifiers were chosen – Support Vector Machine (SVM), Gradient Boosting (GB) and Feedforward Neural Network (FNN). The performance evaluation was done using accuracy, specificity, sensitivity, F1 score, ROC curves, confusion matrix and in the case of FNN the loss function. Ten different model settings were performer for each biosignal, whose stability was verified by calculating the percentage differences in metrics between the training and test/validation sets. Based on the sum of the absolute differences of the means and medians of the metrics, the best classifier was chosen for each biosignal. The thesis also includes an analysis of performance differences after training and after testing/validation, which was used to evaluate the best and worst settings for each classifier. The results showed that the Feedforward neural network (FNN) classifier performed the most stable for the GSR/EDA and RR signals, while the Gradient Boosting (GB) classifier was the most suitable for the HR signal. On the other hand, the Support Vector machine (SVM) method was evaluated as the least suitable classifier.