Central apnea detection in premature infants using machine learning

BACKGROUND AND OBJECTIVE: Apnea of prematurity is one of the most common diagnosis in neonatal intensive care units. Apneas can be classified as central, obstructive or mixed. According to the current international standards, minimal fluctuations or absence of fluctuations in the chest impedance (CI) suggest a central apnea (CA). However, automatic detection of reduced CI fluctuations leads to a high number of central apnea-suspected events (CASEs), the majority being false alarms. We aim to improve automatic detection of CAs by using machine learning to optimize detection of CAs among CASEs. METHODS: Using an optimized algorithm for automated detection, all CASEs were detected in a population of 10 premature infants developing late-onset sepsis and 10 age-matched control patients. CASEs were inspected by two clinical experts and annotated as CAs or rejections in two rounds of annotations. A total of 47 features were extracted from the ECG, CI and oxygen saturation signals considering four 30 s-long moving windows, from 30 s before to 15 s after the onset of each CASE, using a moving step size of 5 s. Consecutively, new CA detection models were developed based on logistic regression with elastic net penalty, random forest and support vector machines. Performance was evaluated using both leave-one-patient-out and 10-fold cross-validation considering the mean area under the receiver-operating-characteristic curve (AUROC). RESULTS: The CA detection model based on logistic regression with elastic net penalty returned the highest mean AUROC when features extracted from all four time windows were included, both using leave-one-patient-out and 10-fold cross-validation (mean AUROC of 0.88 and 0.90, respectively). Feature relevance was found to be the highest for features derived from the CI. A threshold for the false positive rate in the mean receiver-operating-characteristic curve equal to 0.3 led to a high percentage of correct detections for all CAs (78.2%) and even higher for CAs followed by a bradycardia (93.4%) and CAs followed by both a bradycardia and a desaturation (95.2%), which are more critical for the well-being of premature infants. CONCLUSIONS: Models based on machine learning can lead to improved CA detection with fewer false alarms