Interpretable machine learning model for comparing and validating three diagnostic criteria for bronchopulmonary dysplasia in predicting value of respiratory prognosis of preterm infants:a retrospective cohort study

QiQi BuXin WangYanyan wuYingyuan WangRui LiYanmei Zhaocaijun WangGuiying SunWenqing Kang^*

• Zhengzhou Children's Hospital, Zhengzhou, China

Background:Comparison and validation of the predictive value of three diagnostic criteria for bronchopulmonary dysplasia for respiratory prognosis of preterm infants with gestational age <32 weeks. Methods:This retrospective cohort study was conducted to collect clinical data of 397 preterm infants. On the basis of the follow-up results, the enrolled population was divided into a respiratory adverse outcome group and a normal outcome group.The 2001NICHD,the 2018 NICHD , and the 2019 NRN criteria were used to diagnose and grade BPD in preterm infants.The dataset was randomly divided, with 70% used for model training and 30% used for model validation.The extreme gradient boosting machine learning algorithm was used for model training. Furthermore, the SHapley additive exPlanation analysis method was utilized to visually interpret the results of the machine learning model. Results:A total of 397 preterm infants were included. In the training set, prediction models based on the 2001 NICHD,2018NICHD,and 2019NRNcriteria achieved AUC values of 0.747,0.804,and 0.789, with corresponding accuracies of 0.740,0.765,and 0.765. In the test set, the respective AUC values were 0.694,0.747,and 0.752, and accuracies were 0.750,0.800,and 0.750. Based on the DeLong's method, comparisons of ROC curves between the training and test sets revealed that both the 2018 NICHD and 2019 NRN criteria demonstrated significantly higher AUC than the 2001 NICHD criteria (training set: Z=-3.514, -2.110, both P<0.05; test set: Z=-2.137, -2.199, both P<0.05). However, there was no statistically significant difference in the AUC between the 2018 NICHD and 2019 NRN criteria for either the training set (Z=0.863, P=0.388) or the test set (Z=-0.176, P=0.861). The SHAP revealing that the two most important features affecting the respiratory prognosis of preterm infants were the severity of BPD and early invasive ventilation. Conclusions:Both the 2018 NICHD and 2019 NRN criteria for BPD show better and similar predictive values for respiratory adverse outcomes in preterm infants, and both are superior to the 2001 NICHD criteria. The top two factors affecting the respiratory prognosis of preterm infants are the severity of BPD and early invasive mechanical ventilation.