Automation in sleep state classification and maturity evaluation of EEG in preterm-born infants

There is a growing need to automate EEG monitoring of preterm-born infants, particularly where expert evaluation is limited. However, existing models necessitate human intervention for comprehensive assessment, thereby limiting the clinical applicability. We collected 165 EEG records from 65 preterm-born infants. 8-channel EEG records were segmented every 30second and allocated into eight classes (three preterm-equivalent, four term-equivalent, and one noise class). We performed noise detection, sleep state classification, and maturity assessment through the classification. ImageNet pretrained ResNet-50 was finetuned to classify the segmented EEG images. A novel index, named "maturity rate," was calculated as the proportion of term-equivalent labels in each EEG recording. The model achieved a macro F1 score of 0.58 and Cohen's Kappa of 0.58 for the 8-class classification. When the labels were merged into active sleep, quiet sleep, and noise categories, performance improved to an F1 score of 0.84 and Kappa of 0.74. The maturity rate increased continuously with postmenstrual age at the time of recording, reflecting functional brain development beyond the term-preterm dichotomy. Our model enables end-to-end EEG assessment of sleeping newborns by simultaneously performing noise detection, sleep state classification, and maturity assessment, with minimal need for human intervention.