A Data-Driven Multidimensional Framework for Physiological Emotion Characterisation in Immersive Virtual Reality

Objective. Current affective computing relies on supervised models constrained by predefined frameworks such as the valence–arousal circumplex, precluding discovery of intrinsic physiological structures and obscuring inter-individual variability in emotional response. This study investigates the physiological organization of emotional states through an unsupervised, data-driven framework applied to multimodal physiological recordings acquired in virtual reality (VR). Approach. Thirty-eight healthy participants underwent a 35-minute VR protocol inducing seven emotions —Relaxation, Happiness, Sadness, Fear, Disgust, Empathy, and Frustration — via custom immersive scenarios. Simultaneous recordings of electroencephalographic, electrocardiographic, blood volume pulse, galvanic skin response, and respiration signals were acquired; 74 features were extracted and reduced to 36 via correlation-based selection. Principal Component Analysis (PCA) identified latent physiological dimensions, and K-means clustering was applied within each emotion to characterise inter-individual variability. Main Results. PCA revealed four physiologically interpretable dimensions, collectively explaining 40.1% of total variance, capturing distinct autonomic, cardiovascular, and neural signatures of emotional responding. Subjective arousal was broadly reflected in one component, but no dimension mapped onto subjective valence: emotions with opposing valences clustered indistinguishably in physiological space. Exploratory within-emotion clustering revealed marked inter-individual variability in the physiological response to the same emotion. A three-feature multimodal triplet (posterior alpha power, electrodermal phasic reactivity, parasympathetic cardiac tone) achieved complete separation across all emotions. Significance. Unsupervised analysis reveals organizational structures extending beyond standard affective frameworks. Individual physiological response profiles challenge the assumption of homogeneous emotional responses, providing a foundation for label-free, personalised affective computing in clinical and human-computer interaction contexts.