Early prediction of intraventricular hemorrhage in very low birth weight infants using deep neural networks with attention in low-resource settings

Abstract

Early prediction of intraventricular hemorrhage (IVH) in very low-birthweight infants (VLBWIs) remains challenging because of multifactorial risk factors. IVH often occurs within a few hours after birth, yet its onset cannot be reliably predicted using clinical symptoms or vital signs such as blood pressure or heart rate. Accurate early prediction of IVH is critical for timely intervention but remains challenging due to the limited feasibility of current prediction methods in resource-constrained settings. Traditional prediction methods often require advanced equipment and specialized expertise. Therefore, developing a predictive model based on a limited set of available factors is crucial for accurate and reliable early IVH prediction in VLBWIs. We propose a deep neural network-based model with an attention mechanism (DNN-A) that utilizes a limited set of readily available clinical factors. The model was trained and tested on data from 387 infants, incorporating eight variables, including maternal age, delivery mode, endotracheal intubation, birth weight, gestational age at delivery, APGAR scores at 1 and 5 min, and sex. Our DNN-A model achieved 90% and 87% accuracies in the training and testing sets, respectively, with an area under the curve of 87, and outperformed various state-of-the-art machine-learning-based models for IVH prediction. These results underscore the effectiveness of DNN-A for predicting the risk of IVH in VLBWIs in low-resource settings.