Abstract
Introduction
Affective experience emerges from interactions among cortical activity, bodily physiology, and contextual interpretation, yet the mechanisms linking these processes remain incompletely understood. Although EEG-based studies have advanced affect decoding, many data-driven approaches provide limited mechanistic insight into inter-individual variability. We developed a subject-specific, hypothesis-generating computational framework linking EEG-derived cortical activity and cardiorespiratory dynamics to subjective affect through putative neuromodulatory processes.
Methods
The framework integrates a respiratory–circulatory subsystem with spiking-neuron models representing the locus coeruleus, ventral tegmental area, raphe nucleus, and basolateral amygdala. EEG band-power features were used as hypothesis-driven, indirect inputs to these modeled processes rather than as direct measures of deep-brain activity, while cardiorespiratory variables provided physiological feedback to the neural circuit. Model-derived activity was converted into putative norepinephrine-, dopamine-, and serotonin-related signals. Together with two fitted serotonin-related modulation coefficients, these signals constituted five latent parameters used to represent valence–arousal states. The modulation coefficients were estimated for each participant–stimulus observation by nonlinear optimization so that model-derived arousal and valence matched the corresponding subjective ratings. Model calibration and predictive validation were evaluated separately. Predictive validity was assessed using leave-one-stimulus-out cross-validation, in which each stimulus was held out in turn and predicted from regression equations fitted only to the remaining 15 stimuli.
Results
The framework was evaluated using the AMIGOS dataset (nine participants, 16 video stimuli; 144 participant–stimulus observations). Under the calibration procedure, the five latent parameters reproduced participant-specific valence–arousal ratings. In cross-validation across all participants, the mean root mean square error was 0.380 ± 0.288 for arousal and 1.283 ± 1.512 for valence, indicating more consistent prediction of unseen stimuli for arousal and substantial inter-individual variability in valence prediction. Cross-participant regression showed weaker generalization, further supporting the importance of subject-specific modeling.
Discussion
The proposed framework does not directly measure deep-brain or monoaminergic activity. Rather, it provides a physiologically informed and interpretable framework for generating hypotheses about how cortical activity, cardiorespiratory feedback, and putative neuromodulatory processes may jointly contribute to individual affective experience. The findings also highlight the need for validation in larger independent datasets using direct neurophysiological measurements.