Artificial Intelligence for Detecting Pain in Neonates in the NICU: A Systematic Review of the Literature

Background: Accurate neonatal pain assessment remains challenging because behavioural responses are subtle, physiological signals are nonspecific, and commonly used bedside scales are intermittent and observer-dependent. Artificial intelligence (AI) may enable continuous, objective assessment by integrating facial, vocal, behavioural, physiological and neurophysiological information.

Methods: We conducted an updated systematic evidence audit in accordance with PRISMA 2020. PubMed/MEDLINE, IEEE Xplore, ACM Digital Library, publisher and index records, and backward and forward citation searches were examined. Peer-reviewed primary studies evaluating AI, machine-learning or deep-learning approaches for neonatal pain and reporting quantitative performance were eligible. Reviews, preprints, synthetic-only studies without validation in real neonates, and studies addressing outcomes other than pain were excluded. Risk-of-bias domains and reporting completeness were evaluated using PROBAST- and TRIPOD-oriented criteria. Reports based on shared cohorts or benchmark datasets were linked to avoid treating repeated model publications as independent clinical replications.

Results: Seventy unique records were screened, 62 reports were sought for retrieval, 60 full texts were assessed, and 55 primary reports were included, representing 28 distinct study or dataset families. Thirteen included reports were published in 2026. Vision-centred methods predominated (31/55; 56.4%), followed by multimodal systems (16/55; 29.1%), audio-centred approaches (5/55; 9.1%) and EEG-based models (3/55; 5.5%). Recent studies included larger cohorts, multimodal and neuro-symbolic methods, and prospective NICU validation. However, comparison across studies was limited by heterogeneous pain definitions, reference standards, datasets and performance metrics. Recurrent concerns included single-centre samples, class imbalance, dataset reuse, potential split leakage and limited external validation.

Conclusion: AI shows promise as a supportive tool for neonatal pain recognition but is not ready to replace clinical judgement. Future research should prioritize prospective multicentre validation, transparent reporting, external testing, fairness assessment and evaluation of clinical workflow integration.