Explainable machine learning for central apnea detection in premature infants

Diagnosis for apnea of prematurity is commonly performed by detecting central apneas (CAs) in the respiratory traces of premature infants. Previous studies reported that up to 65% of CA alarms sounding in clinical practice are false. We recently showed that using a CA detection model based on elastic net logistic regression (ENLR) and features derived from physiological signals can lead to improved precision. This study explores the possibility of using other explainable machine learning algorithms for the same purpose. CA detection models based on Support Vector Machines (SVM), eXtreme Gradient Boosting (XGBoost) and K-Nearest Neighbors (KNN) were therefore developed using the same dataset consisting of 10 premature infants and leave-one-patient-out cross-validation. Among the new additions, XGBoost led to the development of the most promising CA detection model. It returned a slightly lower mean area under the receiver operating characteristic curve (AUROC) value (i.e., 0.84 vs. 0.86) but also fewer false CA alarms per patient per hour in stable periods located far away from apneic events (i.e., 2.42 vs. 2.58). This result could reduce the burden on nurses in clinical practice, avoiding unnecessary responses during periods when they are least needed. Most features within the first decision trees were consistently selected, regardless of the patient left in the test set. These were also ranked high in terms of feature importance for both the CA detection model based on XGBoost and ENLR, proving their capability to effectively distinguish CAs from stable periods.