Prediction of neonatal death in pregnant women in an intensive care unit: Application of machine learning models

Marcos Espinola-Sánchez; Silvia Sanca-Valeriano; Andres Campaña-Acuña; José Caballero-Alvarado

doi:10.1016/j.heliyon.2023.e20693

Prediction of neonatal death in pregnant women in an intensive care unit: Application of machine learning models

Heliyon. 2023 Oct 5;9(10):e20693. doi: 10.1016/j.heliyon.2023.e20693. eCollection 2023 Oct.

Authors

Marcos Espinola-Sánchez^{1

2}, Silvia Sanca-Valeriano², Andres Campaña-Acuña^{2

3}, José Caballero-Alvarado⁴

Affiliations

¹ Facultad de Ciencias de la Salud, Universidad Privada del Norte, Peru.
² Centro de Innovación e Investigación Traslacional en Salud, Universidad Privada del Norte, Peru.
³ Instituto Nacional Materno Perinatal, Peru.
⁴ Hospital Regional Docente de Trujillo, Peru.

Abstract

Introduction: Neonatal mortality remains a critical concern, particularly in developing countries. The advent of machine learning offers a promising avenue for predicting the survival of at-risk neonates. Further research is required to effectively deploy this approach within distinct clinical contexts. Objective: This study aimed to assess the applicability of machine learning models in predicting neonatal mortality, drawing from maternal and clinical characteristics of pregnant women within an intensive care unit (ICU).

Methods: Conducted as an observational cross-sectional study, the research enrolled pregnant women receiving care in a level III national hospital's ICU in Peru. Detailed data encompassing maternal diagnosis, maternal characteristics, obstetric characteristics, and newborn outcomes (survival or demise) were meticulously collected. Employing machine learning, predictive models were developed for neonatal mortality. Estimations of beta coefficients in the training dataset informed the model application to the validation dataset.

Results: A cohort of 280 pregnant women in the ICU were included in this study. The Gradient Boosting approach was selected following rigorous experimentation with diverse model types due to its superior F1-score, ROC curve performance, computational efficiency, and learning rate. The final model incorporated variables deemed pertinent to its efficacy, including gestational age, eclampsia, kidney infection, maternal age, previous placenta complications accompanied by hemorrhage, severe preeclampsia, number of prenatal checkups, and history of miscarriages. By incorporating optimized hyperparameter values, the model exhibited an impressive area under the curve (AUC) of 0.98 (95 % CI: 0.95-1), along with a sensitivity of 0.98 (95 % CI: 0.94-1) and specificity of 0.98 (95 % CI: 0.93-1).

Conclusion: The findings underscore the utility of machine learning models, specifically Gradient Boosting, in foreseeing neonatal mortality among pregnant women admitted to the ICU, even when confronted with maternal morbidities. This insight can enhance clinical decision-making and ultimately reduce neonatal mortality rates.

Keywords: Artificial intelligence; Intensive care units; Machine learning; Mortality; Neonate; Pregnancy complication.