Adaptive Nonlinear Causal Quantification for Real-Time Emotion Analysis using EMD-based Causal Decomposition
Weifeng Li, Wenbin Shi, Chien-Hung Yeh · 2024
Affective computing, which combines neuroscience and signal processing, plays a critical role in accurately deciphering human emotional states—a key component in various applications. Traditional emotion detection methods often face limitations due to their vulnerability to social influences and reliance on linear models. Electroencephalogram (EEG) signals present a more objective measure by directly reflecting brain activity related to emotional states. This study introduces the EMD-based causal decomposition (CD-EMD) to infer neural causality, which involves decomposing EEGs into intrinsic mode functions (IMFs) and capturing instantaneous phase interactions among IMFs. Unlike the traditional Granger causality, which assumes linearity and stationarity, CD-EMD offers an adaptive and nonlinear framework that more accurately represents the complex, time-varying nature of brain dynamics. Based on the DEAP dataset, the graph metrics across various frequency bands are extracted, after CD-EMD constructs the EEG network by quantifying the causal strength (CS) between each two channels. The results show that positive emotions are associated with increased α band centrality and clustering, suggesting enhanced cognitive functions, while negative emotions are linked to heightened β band clustering and longer path lengths, indicating increased stress responses. This research not solely offers new perspectives on the neural dynamics of emotional states, but also demonstrates the superior time efficiency of CD-EMD. Our findings underline the potential of the proposed CD-EMD in enhancing real-time emotion detection and the untangling of emotional brain networks.