RST-Guided Adaptive Multimodal Fusion Transformer for Interpretable Infant Pain Assessment

Pain assessment in non-communicative infants remains a critical challenge in neonatal intensive care units, where subjective observational scales suffer from inter-rater variability and lack continuous monitoring capabilities. This paper presents RST-AMFT, a novel multimodal fusion framework that integrates Rough Set Theory (RST) for interpretable feature selection with adaptive transformer-based architectures for robust pain intensity estimation. RST is not introduced here for the first time; rather, this work is the first to integrate RST into an end-to-end multimodal temporal fusion framework for infant pain assessment. Our approach dynamically weights four complementary modalities-facial expressions, eye-tracking metrics, physiological signals, and cry audio-based on their real-time reliability at each timestep, addressing challenges of temporal misalignment and missing data. Validated on a multimodal dataset of 133 infants (28-42 weeks gestational age, two clinical sites) with pain labels derived from three established clinical scales (NFCS, COMFORT, NIPS), RST-AMFT achieves 99.5% frame-level accuracy with substantial agreement to clinical scales (Cohen's $\kappa = 0.92$), outperforming all baseline methods. The RST-based feature selection reduces dimensionality by 44% (18 to 10 features) while maintaining interpretability through 47 stable, human-readable decision rules (mean confidence 0.91). Cross-site validation yields 96.8% accuracy ($\kappa =0.88$). Clinician usability ratings of extracted rules reached 4.6/5.0, supporting potential clinical adoption.