MSAFNet: A multi-scale attention fusion network for automated neonatal pain facial expression recognition with...

Accurate, timely pain assessment in neonates remains a stubborn yet critical challenge in neonatal intensive care units (NICUs), where conventional observational scales suffer from subjectivity, intermittent use, and uneven inter-rater reliability. We propose a Multi-Scale Attention Fusion Network (MSAFNet) for automated neonatal pain facial expression recognition, and validate it in an authentic clinical setting. Drawing on two public collections together with a prospectively gathered NICU corpus, we assembled 9,485 annotated facial images from 153 neonates; after strict subject-level partitioning, training-set-only augmentation (MixUp, CutOut, plus conventional geometric and photometric perturbations) expanded the training pool to over 25,000 samples while the validation and test sets remained un-augmented. The MSAFNet architecture combines a modified lightweight ResNet-34 backbone with landmark-guided local feature branches, multi-scale parallel dilated convolution pathways, channel-spatial dual attention, and an adaptive softmax-weighted fusion mechanism, targeting both subtle muscle contractions and holistic configurational change in neonatal faces. On the held-out test set MSAFNet reached 94.3% accuracy, 94.1% precision, 94.1% recall, 94.2% F1-score, and an AUC of 0.978, outperforming ResNet-50, EfficientNet-B0, and domain-specific baselines. A prospective clinical validation covering 62 neonates and 287 procedural pain episodes yielded substantial agreement with expert nurse consensus (Cohen's Kappa = 0.84, ICC = 0.91), with a per-frame inference latency of approximately 0.3 s. Usability ratings from nursing staff were favourable, supporting the role of MSAFNet as a supplementary monitoring aid rather than a replacement for clinical judgement. Taken together, the findings suggest that MSAFNet can serve as a practical decision-support tool for continuous, objective neonatal pain assessment in real NICU environments.