Subject-independent emotion recognition in children using multiple instance learning approach on EDA data

Feeling different emotions can produce physiological responses, including changes in cardiovascular activity, perspiration, and facial expression. Although recognizing emotions from physiological signals remains challenging, such signals may provide useful information for affective computing applications. This study presents a class-aware multiple-instance learning framework for identifying emotional states in typically developing children using electrodermal activity (EDA) signals. The framework incorporates class-specific attention mechanisms to learn feature representations for neutral, positive, and negative emotions while accounting for intra-class variability in physiological responses. EDA data were collected from 15 children using the EmotiBit wearable device, and the proposed framework was evaluated using strict leave-one-subject-out cross-validation. The framework achieved an overall accuracy of 81.96%, representing a modest numerical improvement over the MLP baseline; however, this difference was not statistically significant. The class-aware attention mechanism identified temporal signal segments associated with prolonged conductance elevations, brief phasic fluctuations, and patterns contributing differently to predictions across emotional states. These attention weights should be interpreted as model-based indicators rather than direct physiological explanations. Although the findings support the feasibility of interpretable and subject-independent pediatric EDA modeling, the small sample of 15 children limits the strength of the evidence for a broader generalizability. Therefore, validation using larger, more diverse and independent pediatric cohorts is required.