467 - Machine Learning-based Prediction of the Risk for Sepsis in Extremely Preterm Infants

Poster Number: 467
Publication Number: 467.432

Vivek Shukla, University of Alabama at Birmingham, Birmingham, AL, United States; Avinash Singh, Childrens Hospital of Alabama, Birmingham, AL, United States; Arie Nakhmani, University of Alabama at Birmingham, Birmingham, AL, United States; Namasivayam Ambalavanan, University of Alabama School of Medicine, Birmingham, AL, United States; Waldemar A. Carlo, University of Alabama School of Medicine, Birmingham, AL, United States; Colm P. Travers, University of Alabama School of Medicine, Birmingham, AL, United States

Background: Neonatal sepsis results in substantial morbidity and mortality, especially in preterm infants receiving intensive care. The heart rate characteristics index (HRCi) is a tool that measures heart rate variability and improves the early identification of neonatal sepsis. Pain, medications, and respiratory support can affect the predictive accuracy of HRCi.

Objective: Models using routinely available clinical variables and HRCi will accurately (AUC > 0.70) predict neonatal sepsis 3 to 6 hours before blood cultures were initiated.

Design/Methods: Single tertiary center, retrospective, observational study including preterm infants born between 2014-2019, with birth gestational ages between 22^0/7 to 27 ^6/7 weeks with blood culture positive late-onset ( >72 hours after birth) sepsis.Variables, including hourly vital signs, HRCi, respiratory support type and settings, feeding tolerance, blood transfusion, and laboratory results, including electrolytes, blood gas, and glucose, complete blood and differential count, and blood cultures were collected for 96 hours before the blood culture positive time point and were used as predictors. Models were built using 3- and 6-hourly averages (3- and 6-hour scenarios) of the predictor variables and their corresponding entropy. Predictive models included conventional Logistic Regression and machine learning models, including Random Forest, Naïve Bayes, and XGBoost algorithms. The k-fold cross-validation technique was used for evaluating the model performance. Accuracy, precision, recall, F1, and AUC scores were compared to identify the best-performing model. Random Forest Gini feature importance score was used to identify the top predictors by scenario.

Results: A total of 141 infants were included in the study. The birth gestational age (mean ± SD) of the infants was  24^3/7±1^4/7 weeks, and the birth weight was 669 ±185 grams. The random forest model was overall the best-performing model (Table 1). HRCi was the most important predictive variable in both scenarios. Mean airway pressure, pH, HCO3, and saturation, and PCO2, heart rate, respiratory rate, and saturations were the other top predictors for the 3- and 6-hour scenarios, respectively (Table 2). Adding other top predictors substantially improved the sepsis risk predictive accuracy beyond that of HRCi alone (Table 3).

Conclusion(s): Using advanced machine learning–based modeling methods, the current study showed that routinely available clinical and laboratory variables improve the predictive accuracy of HRCi in the 3 to 6 hours before blood culture initiation in preterm infants with late-onset sepsis.