Machine learning classification of EEG responses to pain-related vs non-pain-related stimulus in preterm infants.

INTRODUCTION

Unmanaged pain in preterm infants can lead to long-term developmental consequences. Current pain assessment methods lack specificity, resulting in possible pain mismanagement in Neonatal Intensive Care Units (NICUs). This study explores the application of machine learning (ML) to differentiate between pain-related and non-pain-related cortical activity in preterm infants.

OBJECTIVE

To evaluate the performance of ML models in distinguishing cortical EEG activity during a painful procedure in preterm infants across different postmenstrual ages (PMAs).

METHODS

This observational study was conducted from June 2015 to May 2024 at Mount Sinai Hospital in Toronto, Canada, and University College London Hospital, United Kingdom. EEG data were collected from 72 preterm infants (27 females) during routine heel lance procedures while held in skin-to-skin contact. Infants' gestational ages ranged from 24 to 36 weeks with a mean PMA of 32.87 weeks. Five ML models-XGBoost, support vector machines, Random Forest, Logistic Regression (LR), and convolutional neural networks-distinguished EEG activity pre-heel and post-heel lance.

RESULTS

Model performance was assessed using accuracy and area under the receiver operating characteristic curve (AUC). In the oldest PMA group (≥34 weeks), LR achieved the highest mean accuracy (82%) and AUC (0.90). Similarly, LR achieved the highest mean accuracy (70%) and AUC (0.94) in the middle PMA group (32-33 weeks, 6 days). In the youngest group (<32 weeks), all models except XGBoost performed relatively the same with a mean accuracy of 76% or 77% and a mean AUC of 0.82 or 0.80.

CONCLUSION

Machine learning models demonstrate potential in distinguishing pain-related cortical activity, offering a pathway for improved neonatal pain assessment in NICUs.