Clinical characteristics of bronchopulmonary dysplasia and the risk of sepsis onset prediction via machine learning models

Abstract

Bronchopulmonary dysplasia (BPD), also known as chronic lung disease, is the most common cause of respiratory morbidity in preterm infants. Sepsis plays a significant role in the pathogenesis of BPD, and the systemic inflammatory response caused by sepsis is associated with lung development, leading to simplified alveoli and abnormal vascular development, which are the histological hallmarks of BPD. In this study, we conducted a retrospective analysis of the clinical characteristics of 306 preterm infants with BPD treated at our hospital from December 2019 to December 2022. We subsequently utilized ten machine learning (ML) algorithms and used clinical features to acquire models to predict BPD with sepsis. The performance of the model was evaluated according to the mean area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. The mean area under the curve (AUC) of the best predictive model was 0.93. A nomogram for sepsis onset was developed in the primary cohort with four factors: invasive respiratory support, CRIB II score, NEC, and chorioamnionitis. By including clinical features, ML algorithms can predict BPD with sepsis, and the random forest (RF) model (sorted by the mean AUC) performs the best. Our prediction model and nomogram can help clinicians make early diagnoses and formulate better treatment plans for preterm infants with BPD.

Keywords: bronchopulmonary dysplasia; machine learning algorithms; nomogram; prediction model; sepsis.

Figure 1

Distribution of birth weight (A) and gestational age (B) in sepsis and nonsepsis BPD patients. BPD, bronchopulmonary dysplasia.

Figure 2

Ml and prediction models for sepsis-induced BPD patients. (A) ROC curves after internal validation via bootstrap resampling (1,000 times) of 10 machine learning models. the shading represents the mean AUC of the bootstrap samples, and the line represents the apparent AUC. (B) SHAP heatmap generated via the random forest model. (C) A nomogram was used to predict sepsis in infants with BPD. A binary logistic regression algorithm was used to establish the nomogram. The final score was calculated as the sum of the individual scores for each of the four variables included in the nomogram. (D) Calibration curve of the regression model. The X-axis represents the overall predicted probability of sepsis in infants with BPD, and the Y-axis represents the actual probability. Model calibration is indicated by the degree of fitting of the curve and the diagonal. (E) DCA curve of the logistic regression model. The horizontal axis in the figure represents the threshold probability, and the vertical axis represents the net benefit (NB). The lines' None “and” All “represent two extreme situations, where” None' indicates that all patients have a negative outcome, no intervention has been performed, and NB is 0. All the lines indicate that all patients have a positive outcome and that all have received intervention. Its NB is a negative sloping diagonal line. In this analysis, the decision curve provided a larger net benefit across the range of 0.2–0.80. (F) Clinical impact curve of the logistic regression model. LR, logistic regression; RF, random forest; SVM, support vector machine; DTREE, decision tree; ADB, AdaBoost; NB, Gaussian naive Bayes; LDA, linear discriminant analysis; KNN, k-nearest neighbors; GB, gradient boosting classifier; MLP, multilayer perceptron.