Abstract

Previous research showed that supervised dimensionality reduction using Neighborhood Components Analysis (NCA) enhanced the performance of 3-class problem emotion recognition using ECG only where features were the statistical distribution of dominant frequencies and the first differences after applying bivariate empirical mode decomposition (BEMD). This paper explores how much NCA enhances emotion recognition using ECG-derived features, esp. standard HRV features with two difference normalization methods and statistical distribution of instantaneous frequencies and the first differences calculated using Hilbert-Huang Transform (HHT) after empirical mode decomposition (EMD) and BEMD. Results with the MAHNOB-HCI database were validated using subject-dependent and subject-independent scenarios with kNN as classifier for 3-class problem in valence and arousal. A t-test was used to assess the results with significance level 0.05. Results show that NCA enhances the performance up to 74% from the implementation without NCA with p-values close to zero in most cases. Different feature extraction methods offered different performance levels in the baseline but the NCA enhanced them such that the performances were close to each other. In most experiments use of combined standardized and normalized HRV-based features improved performance. Using NCA on this database improved the standard deviation significantly for HRV-based features under subject-independent scenario.