Developing and Validating a Multimodal Dataset for Neonatal Pain Assessment to Improve AI Algorithms With Clinical...

Area of Science:

• Neonatal care
• Artificial Intelligence in Medicine
• Pain Management

Background:

• Accurate data labeling is essential for effective Artificial Intelligence (AI) in neonatal pain assessment.
• Reliable neonatal pain datasets are critical for managing pain in newborns.

Purpose of the Study:

• To develop and validate a comprehensive multimodal dataset with accurately labeled clinical data.
• To enhance AI algorithms for neonatal pain assessment through improved data.

Main Methods:

• Healthy neonates were assessed using the Neonatal Pain, Agitation, and Sedation Scale.
• Pain reactions were recorded using two cameras during painful procedures.
• Assessors labeled processed pain data on the EasyDL platform (modality-data neonatal pain labeling - MD-NPL) after two weeks.

Main Results:

• High intraclass correlation coefficients (0.938-0.969) were found among single-modal data.
• An overall pain intraclass correlation coefficient score of 0.99 and a Kappa statistic of 0.899 for pain grade agreement were achieved.
• Linear regression models showed a goodness-of-fit greater than 0.96 when comparing on-site neonatal pain assessment (OS-NPA) and MD-NPL.

Conclusions:

• Modality-data neonatal pain labeling (MD-NPL) is a productive alternative to on-site neonatal pain assessment (OS-NPA).
• The data labels within the Multimodality Dataset for Neonatal Acute Pain are validated.
• These findings provide reliable validation for AI algorithms designed for neonatal pain assessment.