Automatic Apnea Detection and Classification in Preterm Infants Combining Machine Learning with Transcutaneous Electromyography of the Diaphragm: a Pilot Study.

Abstract

Introduction: Real-time monitoring of apneic events is important to improve treatment and reduce intermittent hypoxemia in preterm infants, but remains challenging. We aimed to develop and test a machine learning (ML) model capable of automatic detection and classification of apneic events in preterm infants, using transcutaneous electromyography of the diaphragm (dEMG) and clinical parameters.

Methods: In this prospective pilot study preterm infants (26-32 weeks gestational age) were monitored for 4-8 hours, recording dEMG, airway pressure (AWP), and oxygen saturation. Afterwards, experts labelled apneic events as central apnea (CA), obstructive apnea (OA), or signal noise. 60% of the data was used for model training, the remaining data was split into a validation and test dataset. Main outcomes were model sensitivity and positive predictive value (PPV).

Results: Thirty infants with a median birth gestational age of 27.6 (IQR 26.9-28.4) weeks and birthweight of 990 (IQR 878-1177) grams were included. 21 infants were used for modelling, as AWP data turned out to be only reliable during nasal CPAP. In total, 89.5 hours of data, containing 1093 apnea labels were used. In the test set, which contained 76 CA and 141 OA labels, the ML model detected 53 and 103 events, respectively. This resulted in sensitivity and PPV scores of 70% and 88% for CA events, and 73% and 64% for OA events.

Conclusion: A dEMG-based machine learning model is feasible for automatic detection and classification of apneic events in preterm infants. Future research will focus on model validation in a larger cohort.