A Pilot Study Using Machine Learning for Classification of Pain-Related Versus Non-Pain-Related Electroencephalographic Activity in Preterm Infants

Effective pain assessment and management are crucial to mitigate both immediate and long-term consequences of prolonged NICU stays. Accurately assessing pain in premature infants is challenging due to their inability to verbally communicate their pain, the potential judgement bias by caregivers, the lack of specificity in current pain assessment tools and time constraints in a busy hospital environment. This pilot study explores a machine learning approach to support pain assessment in neonatal care using cortical activity.

The current study aims to test machine learning models that autonomously distinguishes non-pain related from pain-related cortical activity. The present dataset includes 72 preterm infants (27 females), born between 24- and 36-weeks gestational age, from two NICUs: Mount Sinai Hospital (Toronto, Canada) and University College London Hospital (London, UK). The primary outcome was to assess the accuracy of various machine learning models (XGBoost, Support Vector Machines, Random Forest, Logistic Regression, Convolutional Neural Networks) in distinguishing EEG features within a one-second pre-lance epoch (non-pain related) from a one-second post-lance epoch (pain-related). Performance metrics varied across post-menstrual age groups, reflecting developmental differences in EEG patterns. Machine learning algorithms can autonomously distinguish the one-second epoch immediately following a heel lance from the one-second epoch immediately preceding the procedure in preterm infants. Moreover, the performance of these algorithms improves with increasing postmenstrual age, demonstrating greater accuracy and reliability in older infants. This study provides a foundation for developing an autonomous and accurate tool for pain assessment in neonatal patients that can improve pain management practices in NICUs.