Estimating infant mortality risk in Uruguay using artificial neural networks

Abstract

Abstract Background Uruguay has a national electronic live birth certificate that includes variables of risk factors for infant mortality. Accurate risk stratification of newborns is needed to optimize the use of resources for homes visits. Artificial neural networks are computational tools that have been used successfully in many types of prediction problems. Objective To develop a neural network able to estimate the risk of infant mortality using information available in the electronic live birth and mortality certificate in Uruguay. Methods A historical cohort of records of newborns in Uruguay from 2014 to 2017 was used. The variables of the electronic live birth certificate were considered for the model. Infant mortality was obtained from the national mortality registry. A multilayer perceptron was trained with a random sample of 70% of the cohort; the remaining 30% was the validation set. The variables included were birth weight, Apgar score at 5 minutes, number of prenatal consultations, maternal educational level, multiple pregnancy and cohabiting father. ROC curve analysis was performed. Results The 2014-2017 birth cohort contains 187,388 records. 1,307 children under one year died (IMR 6.97 per 1,000 births). The area under the curve (AUC) was 88.7%, 95% CI [87.6% - 89.8%]. The optimal cut-off point of pseudoprobability of infant mortality was 0.008 (78.9% of sensibility and 82.4% of specificity). The IMR in high-risk newborns identified by the neural network was 32.8 per 1,000 births. Conclusions The neural network identifies high-risk newborns at the time of entering the data in the electronic live birth certificate as other models have done. This information could be used to plan and implement preventive actions. Key messages In Uruguay, high-risk newborns can be identified by applying artificial intelligence to data collected routinely. The procedure can be applied in other countries with electronic birth certificate.