Prediction of bradycardia in preterm infants using artificial neural networks

Abstract

Bradycardia is common in preterm infants and associated with a range of adverse outcomes, including end organ damage and developmental problems. This paper proposes a method to develop a generalised model to predict the onset of bradycardia in preterm infants by monitoring vital signs using artificial neural networks (ANN). Data used for network development was collected from a study conducted at the Royal Hobart Hospital involving 31 preterm infants, and comprising 3591 h of electrocardiogram (ECG) and respiratory motion recordings. ANNs with a multilayer perceptron architecture were employed with features from the ECG and respiratory signals as inputs. The ANN was trained to predict bradycardia within a pre-bradycardia period beginning 15 s prior to each bradycardic event. The ANN’s prediction capability was assessed using the area under the curve (AUC) of the receiver operating characteristic. Heart rate variability and respiration patterns were found to be indicative markers of an impending bradycardic event. When applied to new infants, the ANN using only ECG features achieved a mean AUC of 0.63, and the ANN using both respiratory features and ECG features achieved a mean AUC of 0.69. This approach has improved on previous attempts to predict bradycardia and should be further investigated.