EvoGraphConPain: an adaptive multimodal graph intelligence framework for neonatal pain assessment

Neonatal pain detection remains a critical challenge in NICUs, where subtle, non-verbal cues can lead to under- or over-treatment if not accurately interpreted. Inspired by recent advances in graph-based modeling for inter-modality relationships and contrastive self-supervision for data-efficient learning, we introduce EvoGraphConPain–a novel graph neural network (GNN) framework for continuous infant pain monitoring. The model integrates facial expressions, body movements, cry acoustics, and physiological signals (e.g., heart rate, skin conductance) into a dynamic graph structure, employs contrastive learning for unsupervised pretraining on limited datasets, and uses recurrent layers for temporal dynamics. Unlike prior works focused on transformer-centric fusion or causal invariance, EvoGraphConPain emphasizes graph-learned correlations between modalities and includes a dedicated module for detecting “silent” pain (e.g., without audible cries). Tested on a combined dataset from iCOPE, NPAD, and extended body movement annotations, it achieves 88.5% accuracy in pain level classification (4-level scale), a mean squared error (MSE) of 0.32 for continuous scoring, and 92% recall for silent pain episodes—outperforming the strongest supervised multimodal baseline by 4.4% accuracy and 4.8% silent pain recall (Wilcoxon P < 0.001). With integrated explainability via node importance, EvoGraphConPain supports clinical decision-making for timely interventions, pending prospective clinical validation.