A Feasibility Study on Improving Emotion Recognition from ECG Signals and HRV Features Through Baseline Clusterization

Electrocardiography (ECG) has the potential for bringing Affective Computing outside laboratories, thanks to the spread of wearable and inexpensive instrumentation. Nevertheless, intra individual variability could influence Machine Learning (ML) models’ accuracy. To assess this issue, we propose to group the participants according to their general cardiovascular status, through the clusterization of HRV baseline features. A specific ML model aimed at classifying emotional responses was developed for each baseline cluster. This processing will lead to cardiac-state specific classification models to mitigate ML performance issues. We experimented this data analytics framework on the Mahnob HCI database containing ECG paired with emotional self-report assessment. Baseline data was clustered using k-means, dividing the dataset into two parts. Successively, classification models were separately applied to each group to predict arousal, valence, and dominance levels from ECG features. Classifiers applied after clustering outperformed those without clustering, reaching higher scores and lower randomness. Clustering ECG baselines to create individualized classifiers may alleviate intra-individual variability and improve emotion recognition performance, making affective computing more applicable.