Predictive Model for Motor Developmental Delay in Preterm Infants by Using Recurrent Neural Network

Abstract

Objective: The aim of this study is to develop the predictive model for motor developmental delay in Korean preterm infants beyond neonatal intensive care unit. Methods: The authors retrospectively investigated the medical records of premature infants who had undergone developmental test and discharged from the single regional newborn intensive care center. We collected 30 independent variables and the motor scale of the Korean version of Bayley scale of infant and toddler development III (K-Bayley III). The predictive modeling was conducted by 3 steps: 1) data preprocessing, 2) training predictive models, and 3) evaluation of final performance of each model. We used sensitivity as a primary evaluation metrics, and F1 score and area under precision- recall curve (AUPRC) as a secondary metrics. Results: Total 359 subjects were enrolled in the study. Ten percent of subjects were below 80 in the motor scale (coding as ‘1’ in the dependent variable). Recurrent neural network model showed the best performance (sensitivity 1.00, F1 score 0.36, AUPRC 0.22). XGBoost model (sensitivity 0.71, F1 score 0.63, AUPRC 0.65) and ridge logistic regression model (sensitivity 0.71, F1 score 0.56, AUPRC 0.60) also showed good performance. Conclusion: Machine learning approach showed good predictive value for motor delay in Korean preterm infants. The further research by using big data from multicenter is needed.