Learning-Based Longitudinal Prediction Models for Mortality Risk in Very-Low-Birth-Weight Infants: A Nationwide Cohort Study

Abstract

Introduction: Prediction models assessing the mortality of very-low-birth-weight (VLBW) infants were confined to models using only pre- and perinatal variables. We aimed to construct a prediction model comprising multifactorial clinical events with data obtainable at various time points. Methods: We included 15,790 (including 2,045 in-hospital deaths) VLBW infants born between 2013 and 2020 who were enrolled in the Korean Neonatal Network, a nationwide registry. In total, 53 prenatal and postnatal variables were sequentially added into the three discrete models stratified by hospital days: (1) within 24 h (TL-1d), (2) from day 2 to day 7 after birth (TL-7d), (3) from day 8 after birth to discharge from the neonatal intensive care unit (TL-dc). Each model predicted the mortality of VLBW infants within the affected period. Multilayer perception (MLP)-based network analysis was used for modeling, and ensemble analysis with traditional machine learning (ML) analysis was additionally applied. The performance of models was compared using the area under the receiver operating characteristic curve (AUROC) values. The Shapley method was applied to reveal the contribution of each variable. Results: Overall, the in-hospital mortality was 13.0% (1.2% in TL-1d, 4.1% in TL-7d, and 7.7% in TL-dc). Our MLP-based mortality prediction model combined with ML ensemble analysis had AUROC values of 0.932 (TL-1d), 0.973 (TL-7d), and 0.950 (TL-dc), respectively, outperforming traditional ML analysis in each timeline. Birth weight and gestational age were constant and significant risk factors, whereas the impact of the other variables varied. Conclusion: The findings of the study suggest that our MLP-based models could be applied in predicting in-hospital mortality for high-risk VLBW infants. We highlight that mortality prediction should be customized according to the timing of occurrence.