Abstract
Observer-dependent infant pain scales have limitations including discontinuous assessments and the lack of healthcare professionals' availability. We hypothesized that applying agnostic machine learning approaches to neonatal electroencephalographic (EEG) analysis may reveal features of the infant response to acute pain. EEG was recorded from 30 neonates undergoing acutely painful procedures (18 males, 34.0-41.7 weeks gestation at birth). EEG recordings were randomly assigned to training (n = 20) and testing (n = 10) datasets. Functional connectivity measures were calculated for each infant before and after pain-inducing procedures. A grid search including five machine learning models was conducted on the training dataset, and each model was evaluated using leave-one-subject-out cross-validation. An optimal model, having the highest F-1 score, was obtained and evaluated on the independent testing dataset. A gradient boosting model with 12 features showed optimal performance, with 90% area under the receiver operating characteristic curve suggesting high specificity (0.90) and precision (0.90). The five highest ranked features corresponded to EEG electrode pairs: T7-P4, Fz-CP5, FC1-TP10, CP6-Cz, and Fz-F3, suggesting involvement of the contralateral temporal gyrus, opercular cortex, thalamus, and bilateral insula in infant pain processing. Preliminary changes in functional connectivity indicate infant pain processing. Future machine learning algorithms can integrate physiological and behavioral parameters with EEG changes to accurately assess the complexity of infant pain responses. Trial Registration: ClinicalTrials.gov identifier: NCT03330496.