Introduction: Accurate pain assessment in neonates and infants is essential to prevent adverse neurodevelopmental outcomes but remains challenging due to their inability to self-report. Clinicians rely on behavioral scales such as the Neonatal Infant Pain Scale (NIPS) and Premature Infant Pain Profile-Revised (PIPP-R), which exhibit high inter-rater variability and contextual bias. This work introduces a clinically validated multimodal deep learning framework for objective, continuous, real-time pain intensity estimation in neonatal intensive care settings.
Methods: The proposed architecture integrates three modality-specific encoders: an Inflated 3D-ResNet-50 pretrained on Kinetics-400 for spatiotemporal analysis of 16-frame facial video clips, a VGGish backbone with transformer encoder for log-Mel spectrograms of cry segments, and a 1D-CNN for synchronized 30-second windows of heart rate variability (HRV) and oxygen saturation (SpO₂). Multi-head cross-modal attention dynamically aligns and fuses these representations before regressing a continuous pain intensity score, calibrated to the NIPS 0–10 scale. The framework represents a validated research prototype that has completed prospective clinical evaluation and is in the pre-implementation phase, requiring multi-site validation before routine clinical deployment.
Results: Prospectively evaluated on 127 infants (gestational age 28–42 weeks) undergoing routine painful procedures across two neonatal intensive care units, the system achieved a mean absolute error (MAE) of 0.84 and an intraclass correlation coefficient (ICC) of 0.93 against expert-rated NIPS scores from two senior neonatologists (inter-rater ICC = 0.95). The method significantly outperformed all unimodal and early-fusion baselines (Wilcoxon signed-rank test, p < 0.001). Inference latency was below 180 ms on a single GPU, demonstrating real-time feasibility.
Conclusions: This framework offers real-time, objective, expert-level pain intensity estimation with strong clinical potential for reducing observer bias and enabling evidence-based personalized analgesia in neonatal intensive care.
Current clinical status: This is a validated research tool demonstrating equivalence to expert assessment in controlled settings, with ongoing multi-center trials and clinical workflow integration protocols being under development.
