Automated estimation of EEG maturity in preterm neonates and its association with long-term outcome.

OBJECTIVE

To develop convolutional neural network (CNN) models to estimate EEG maturational age (EMA) from EEG recorded via amplitude-integrated EEG monitors in preterm infants and evaluate its association with long-term neurodevelopmental outcomes.

METHODS

Three CNN models were trained on EEG data recorded from aEEG monitors from 92 preterm infants (23-41 weeks postmenstrual age) without major neurological complications and with normal cognitive outcomes at 2 years. The best-performing model based on MAE was applied to a broader cohort of 148 infants to assess the relationship between predicted age difference (PAD = EMA - PMA) and neurodevelopmental outcome using the Bayley Scales of Infant and Toddler Development (BSID-III).

RESULTS

The best models achieved 87 % and 84 % accuracy within ± 1 week of actual PMA and a mean absolute error of 0.63 and 0.55 weeks. Infants with severely abnormal cognitive outcomes had significantly lower PAD scores compared to those with normal outcomes (P < 0.001). PAD and MAE showed moderate predictive value (AUC 0.69 and 0.77, respectively).

CONCLUSIONS

CNN-based EMA estimation from EEG recorded via aEEG monitors is accurate and correlates with long-term cognitive outcomes in preterm infants.

SIGNIFICANCE: This study demonstrates the clinical potential of automated EEG maturity tracking using EEG as a real-time biomarker for neurodevelopmental risk.