Alasaarela Outline Background Materials and Methods Results and Discussions Conclusions Acknowledgement Background The Mahnob-HCI database provides database for research in affective computing. The baseline accuracy for valence and arousal from ECG signal only using HRV based features are 42.6% and 47.7% respectively. Are there any other features which can increase these accuracies? Materials and Methods The Mahnob-HCI database Recorded from 27 healthy subjects. 32 channels EEG signal Peripheral physiological signals (ECG, temperature, respiration, skin conductance). Face and body videos Synchronized measurement for multimodal emotional response study. Materials and Methods Complex ECG Signals Feature Extractions Each ECG signal is divided into 5 second segment. Complex ECG signal is made based on this segment. Bivariate empirical mode decomposition (BEMD) is applied to the complex ECG signal, resulted 5-6 intrinsic mode functions (IMFs). Only the first three of IMFs are used for further processes. Spectrogram analysis is used to get the dominant frequencies from the IMFs. Feature Extractions Dominant frequency (DF) for each segment is combined with the ones from other segments to represent dominant frequency over the entire ECG signal. Features are statistical distribution of DFs and their first difference. Apply sequential forward floating search algorithm to find the best features combination. Feature Extractions Features are calculated based on the combinations of window size and overlap parameters in spectrogram analysis. It is important to find the optimum combination to get the highest accuracy. As a comparison, features are also calculated from the instantaneous frequencies (IFs) of IMFs after applying empirical mode decomposition (EMD) and BEMD. Validation 10-fold cross validation. 80% of the data is for training and testing using 10-fold cross validation. 20% of the data is for validation. Leave-one-subject-out (LOSO) validation. It is similar to leave-one-out (LOO) but it is applied to all data from one particular subject. The final accuracy is the average accuracy from individuals. Results for 10-fold cross validation Valence Arousal
55.8 7.3 59.7 7.0
Window size = 150 samples Window size = 250 samples Overlap 0 = 30% 1 2 0 1 2 Overlap = 30% 76. 13. 0 73.2 17.9 8.9 0 9.5 6 9 1 44.5 46.1 9.4 38. 44. 16. 56. 1 2 29.9 13.8 8 7 5 3 0:low, 37. 1:medium, 2: high 18. 43. 0:low, 1:medium, 2: high 2 9 5 6 Compare with the others (10-fold cross validation) DF of BEMD Feature (propose IF of IF of calculatio HRV d EMD BEMD n method features ) Classifier SVM KNN KNN KNN Valence 42.6% 55.8% 45.3% 45.6% Arousal 47.7% 59.7% 47.1% 46.6% Results for LOSO validation Valence Arousal
59.2 11.4 58.7 9.1
Window size = 150 samples Window size = 250 samples Overlap 0 = 30% 1 2 0 1 2 Overlap = 30% 81. 11. 0 75.9 18.3 5.8 0 7.4 6 0 1 48.6 43.6 7.8 35. 46. 18. 47. 1 2 40.4 11.9 6 2 2 7 0:low, 37. 1:medium, 2: high 19. 43. 0:low, 1:medium, 2: high 2 4 0 6 Compare with the others (LOSO validation) DF of Feature BEMD IF of IF of calculation (proposed EMD BEMD method features) Classifier KNN KNN KNN Valence 59.2% 45.4% 41.6% Arousal 58.7% 47.0% 46.4% Conclusions The proposed features improve the accuracy to classify emotion in valence and arousal for 3-class problem: from 42.6% to 55.8% for valence and from 47.7% to 59.7% for arousal in 10-fold cross validation (increased by 12%). LOSO validation tests if the proposed features are subject independent Valence: 59.2 11.4 Arousal: 58.7 9.1 Conclusions The proposed features outperformed the other features not only for 10-fold cross validation but also for LOSO validation. Large overlap is not suitable for ECG signal as the frequency contents in P, Q, R, S, and T in the ECG signals have different characteristic. The proposed features may possess potential for building emotion tracking systems because it requires short-time ECG signals. Acknowledgement This research is supported by the Directorate General of Higher Education, Ministry of Higher Education and Research, Republic of Indonesia, No. 2142/E4.4/K/2013 and the Optoelectronics and Measurement Techniques unit, University of Oulu, Finland. The presentation is supported by the IEEE Computational Intelligence Society (IEEE CIS) through Travel Grant award. Thank you