PURPOSE

Predicting individual risks for adverse outcomes in preterm infants is necessary for perinatal management and antenatal counseling for their parents. To evaluate whether a machine learning approach can improve the prediction of severe infant outcomes beyond the performance of conventional logistic models, and to identify maternal and fetal factors that largely contribute to these outcomes.

METHODS

A population-based retrospective study was performed using clinical data of 31,157 infants born at < 32 weeks of gestation and weighing ≤ 1500 g, registered in the Neonatal Research Network of Japan between 2006 and 2015. We developed a conventional logistic model and 6 types of machine learning models based on 12 maternal and fetal factors. Discriminative ability was evaluated using the area under the receiver operating characteristic curves (AUROCs), and the importance of each factor in terms of its contribution to outcomes was evaluated using the SHAP (SHapley Additive exPlanations) value.

RESULTS

The AUROCs of the most discriminative machine learning models were better than those of the conventional models for all outcomes. The AUROCs for in-hospital death and short-term adverse outcomes in the gradient boosting decision tree were significantly higher than those in the conventional model (p = 0.015 and p = 0.002, respectively). The SHAP value analyses showed that gestational age, birth weight, and antenatal corticosteroid treatment were the three most important factors associated with severe infant outcomes.

CONCLUSION

Machine learning models improve the prediction of severe infant outcomes. Moreover, the machine learning approach provides insight into the potential risk factors for severe infant outcomes.