부산대학교 도서관

Hypoxemic events are commonly experienced in preterm infants while in a Neonatal Intensive Care Unit (NICU). These events can contribute to a range of adverse outcomes, including retinopathy, neurodevelopmental impairment, and mortality. Current management of events is focused on recovery rather than prevention. If these events could be predicted, preemptive intervention could be initiated that may act to reduce the duration or prevent hypoxic events altogether. This study aimed to investigate the effectiveness of a machine learning model in predicting hypoxemic events and providing better care than the currently used threshold-based alarms in NICU. Recordings of oxygen saturation, respiratory patterns, and electrocardiogram (ECG) waveforms in preterm infants were extracted from standard cardiorespiratory monitors. A feature analysis was conducted to identify predictors of hypoxemic events, which were used as inputs to train an artificial neural network (ANN) model. A total of 1318 hypoxemic events were used for training and validation of the predictive model. The feature analysis revealed that hypoxemic events were preceded by increased breathing irregularity and greater variability in heart rate (HR), which could assist in predicting their occurrence. The average respiratory rate and the number of prolonged RR intervals (RRIs) in the ECG signal in the periods prior to event onset were identified as potential predictors of the prolonged hypoxemia. In addition, the ANN’s prediction performance outperformed the current threshold-based alarm system in terms of higher accuracy, lower false alarm rate, and longer prediction horizon. The model predicts 56% of hypoxemic events at 30 s prior to the event onset. These findings highlight the potential benefits of utilizing machine learning models in the prediction and prevention of hypoxemic events and could ultimately lead to improved outcomes for preterm infants in the NICU.