Improving the Performance of Stress Detection Systems by Minimizing Person Specific Variations

In this paper, we explore several approaches for improving the performance of a stress detection system (SDS) with electrocardiogram (ECG) signal as its input. Time and frequency domain features extracted from the heart rate variability (HRV), is input to a support vector machine (SVM) backend classifier, for the baseline SVM-SDS. We used ECG data from normative population and patients with severe stress disorders for validating the effectiveness of SDS. We observed that there are variations in the HRV parameters among the normative population, and the patients with stress disorder, usually referred to as person specific variations. To minimize the person specific variations, we explored three methods: 1. Locality-constrained linear coding (LLC) - input feature vectors are expressed in terms of person independent basis vectors. 2. Fisher vector encoding (FVE) -input features are expressed in a person independent feature space using Gaussian mixture models (GMM) trained patient independently. 3. Covariance normalization (CVN) - minimizes the person specific variations in the input vector through feature normalization. For the CVN-SDS, we obtained a sensitivity of 87.50% absolute, specificity of 100.00% absolute and accuracy of 91.66% absolute, which is 31.43%, 31.43% and 33.33% absolute improvement over the SVM-SDS.