Machine Learning-Enabled Emotion Recognition by Multisource Throat Signals

Emotion monitoring plays a crucial role in mental health management. However, traditional methods of emotion recognition predominantly rely on subjective questionnaires or facial expression analyses, which are often inadequate for continuous and highly accurate monitoring. In this study, we propose a high-precision, fine-grained emotion recognition system based on multisource throat physiological signals. The system collects signals through optimized flexible multiporous skin sensors and analyzes them using machine learning models capable of efficiently capturing complex feature interactions. First, we adopt a two-step cross-linking strategy to modulate the porous structure of the sensitive layer to enable accurate detection of the diverse and weak physiological signals in the throat. By extracting four-dimensional features from the input of 7025 samples, the platform based on the Light Gradient Boosting Machine (LightGBM) efficiently captures their nonlinear interactions, ultimately achieving precise classification of five emotional states (relaxation, surprise, disgust, fear, and neutral) with an accuracy of 98.9%. Further validation on an independent data set reveals an average emotion recognition accuracy of 99.3%, demonstrating the system's robustness and reliability in real-world applications. This work provides a viable technological solution for real-time and continuous emotion monitoring, offering significant potential in mental health management and related fields.