Abstract

Emotion recognition systems hold significant practical value due to the vital role emotions play in daily human life. Since cardiac activities are critically involved in the process of emotional arousal, developing emotion recognition systems based on cardiac signals is of great importance. However, inter-subject variability causes a major challenge for cross-subject emotion recognition and remains a key bottleneck for the practical application of emotion recognition systems.

Here we report a photonic cross-subject emotion recognition system (PCERS) based on seismocardiography (SCG) signals, leveraging machine learning techniques driven by complex network feature engineering to decode subject-invariant emotional information from signals. In the cardiac activity monitoring component, we developed a photonic system for SCG signal detection and implemented a sample entropy-based signal processing pipeline. These designs enable precise cardiac activity monitoring as the foundation for emotion recognition.

In the emotion recognition component, the complex network features extracted from SCG signals show significant differences between different emotional states, but no significant differences across subjects. Incorporating these features into the machine learning pipeline enables efficient cross-subject emotion recognition, achieving accuracies 81.25% in leave one-out (LOO) subject-independent emotion recognition. Results in this work suggested that PCERS has potential to contribute meaningfully to practical, real-life emotion recognition applications.