573 - Early Onset Sepsis Risk Prediction in Extremely Preterm Infants

Publication Number: 573.5074

Vivek Shukla, University of Alabama at Birmingham, Birmingham, AL, United States; Avinash Singh, Children's of Alabama, Birmingham, AL, United States; Tao Chen, UAB, Birmingham, AL, United States; Arie Nakhmani, University of Alabama at Birmingham, Birmingham, AL, United States; Colm P. Travers, UAB, Birmingham, AL, United States; Namasivayam Ambalavanan, University of Alabama School of Medicine, Birmingham, AL, United States; Waldemar Carlo, UAB School of Medicine, Birmingham, AL, United States

Background: Predicting the risk of early onset sepsis in extremely preterm infants may reduce unnecessary exposure to antibiotics.

Objective: To develop a prediction model for early-onset sepsis risk in extremely preterm infants.

Design/Methods: We included extremely preterm infants ( < 29 weeks gestation) who received active postnatal treatment, so we had complete data on early onset sepsis. Infants with prenatally diagnosed genetic or metabolic syndromes or major malformations were excluded. The study included infants delivered at the University of Alabama at Birmingham from January 1, 2013–December 31, 2023. All available prenatal and delivery variables were included in the analysis. To address class imbalance, we used a hierarchical Majority Class Removal method. Additionally, we applied the Mutual Information Gain feature selection method to remove irrelevant features, identifying the top 15 important features by measuring how strongly each one relates to the target variable, capturing both simple and complex patterns. Five machine learning models (CatBoost, HistGradientBoosting, Gradient Boosting, XGBoost, Naive Bayes) with class weighting were employed to further improve minority class predictions. Model performance was evaluated using area under the curve (AUC), precision, recall, and F1 score. Feature importance score with incremental AUC analysis was done to identify most important predictors.

Results: Of the 5,777 extremely preterm infants, 77 (1.3%) developed early-onset sepsis. After applying the hierarchical Majority Class Removal method, the dataset was reduced to 2,164 infants (2,087 non-sepsis, 77 sepsis). Infants with sepsis had a significantly lower birth weight (mean ± SD = 773 ± 259 g vs. 913 ± 286 g, p< 0.001, Figure 1) and gestational age (25.7 ± 2.5 weeks vs. 27.3 ± 2.7 weeks, p< 0.001) compared to non-sepsis infants (Table 1). CatBoost emerged as the top-performing model, achieving an AUC of 0.86 and recall of 86.67%, successfully classifying 13 out of 15 sepsis cases in the holdout set. Precision for CatBoost was 11.82%, indicating the challenges in predicting rare events despite high recall (Table 2). CatBoost identified birth weight as the top predictor (feature importance score on a relative scale: 24.67, AUC: 82.57). Other important variables included maternal age (importance score: 15.97), gestational age (importance score:12.88), parity (importance score:11.24), and Apgar 5 score (importance score:10.91).

Conclusion(s): Machine learning models had good performance for predicting early-onset sepsis in extremely preterm infants, with low birth weight emerging as the top predictive variable.

Table 1: Baseline Characteristics


Table 2: Model Performance Comparisons
AUC: area under the curve

Figure 1: Predicted probability of early-onset sepsis as a function of birth weight.
The blue line indicates the predicted probability of sepsis, while the shaded red area represents the 95% confidence interval. The figure demonstrates that the probability of early-onset sepsis decreases as birth weight increases, further supporting the importance of birth weight as an important predictor of sepsis risk in extremely preterm infants.