Performance Evaluation of a Supervised Machine Learning Pain Classification Model Developed by Neonatal Nurses

BACKGROUND: Early-life pain is associated with adverse neurodevelopmental consequences; and current pain assessment practices are discontinuous, inconsistent, and highly dependent on nurses' availability. Furthermore, facial expressions in commonly used pain assessment tools are not associated with brain-based evidence of pain. PURPOSE: To develop and validate a machine learning (ML) model to classify pain. METHODS: In this retrospective validation study, using a human-centered design for Embedded Machine Learning Solutions approach and the Neonatal Facial Coding System (NFCS), 6 experienced neonatal intensive care unit (NICU) nurses labeled data from randomly assigned iCOPEvid (infant Classification Of Pain Expression video) sequences of 49 neonates undergoing heel lance. NFCS is the only observational pain assessment tool associated with brain-based evidence of pain. A standard 70% training and 30% testing split of the data was used to train and test several ML models. NICU nurses' interrater reliability was evaluated, and NICU nurses' area under the receiver operating characteristic curve (AUC) was compared with the ML models' AUC. RESULTS: Nurses weighted mean interrater reliability was 68% (63%-79%) for NFCS tasks, 77.7% (74%-83%) for pain intensity, and 48.6% (15%-59%) for frame and 78.4% (64%-100%) for video pain classification, with AUC of 0.68. The best performing ML model had 97.7% precision, 98% accuracy, 98.5% recall, and AUC of 0.98. IMPLICATIONS FOR PRACTICE AND RESEARCH: The pain classification ML model AUC far exceeded that of NICU nurses for identifying neonatal pain. These findings will inform the development of a continuous, unbiased, brain-based, nurse-in-the-loop Pain Recognition Automated Monitoring System (PRAMS) for neonates and infants.