Development of a Predictive Model for Neonatal Hospital-Acquired Gastrointestinal Infections Utilizing Multiple Machine Learning Algorithms.

Objective: To develop and validate a predictive model for neonatal gastrointestinal infections using multiple machine learning algorithms.

Methods: We conducted a retrospective analysis of 176 neonates diagnosed with nosocomial gastrointestinal infections in NICU between 2020 and 2024, along with a randomly selected control group of 675 neonates without such diagnoses during their NICU stay. The study examined 29 perinatal and NICU treatment-related risk factors potentially associated with neonatal gastrointestinal infections. The dataset was randomly partitioned into training and testing sets. To address class imbalance and enhance minority class identification, we applied SMOTE to the training set.Feature selection used Boruta, Lasso, and Logistic regression, with consensus features from Venn analysis.Subsequently, eight machine learning algorithms were implemented to construct predictive models. Models were evaluated using AUC, F1 score, accuracy, sensitivity, and specificity.

Results: The model incorporated nine significant feature variables: gestational age, NE, PLT, central venous catheterization, nasogastric feeding, delivery mode, intrauterine distress, pregnancy-induced hypertension, and probiotic administration. Among the eight machine learning algorithms evaluated, the Neural Network model demonstrated optimal performance - achieving perfect metrics in the training set (AUC=0.895, F1=0.845, Accuracy=0.837, Sensitivity=0.888, Specificity=0.786, Precision=0.806) and robust results in the test set (AUC= 0.876, F1=0.862, Accuracy=0.856, Sensitivity=0.896, Specificity=0.817, Precision=0.830) - thus was selected as the final predictive model. Model interpretability was enhanced through SHAP analysis. Furthermore, a Shiny-based interactive web calculator for neonatal gastrointestinal infection risk prediction was successfully developed based on this model.

Conclusion: The model effectively identifies at-risk neonates early, supporting clinical decision-making and timely interventions.