Supervised EEG-Based Vigilance State Classification in Preterm Infants

Monitoring sleep in preterm infants is crucial for assessing early brain function, yet most staging tools are designed for term neonates and overlook the unique features of preterm sleep. We developed an automated classifier of vigilance stages tailored to preterm EEG recordings from twenty-two infants (31-36 weeks PMA). Using 1500 30-second epochs of active sleep (AS), quiet sleep (QS), or quiet wakefulness (QW), we first trained a logistic regression with elastic net regularization on time-resolved EEG metrics, achieving 85% accuracy (F1: 84%, Matthews correlation coefficient: 0.75). By incorporating advanced descriptors of brain dynamics, such as phase synchronization, bistability, aperiodic activity, spectral power, and phase-amplitude coupling, the accuracy rose to 97%, highlighting the added value of these constructs. Our model offers a reliable, efficient tool for preterm sleep staging that is well-suited for clinical and resource-limited settings.