Development and Validation of a Machine Learning-Based Risk Prediction Model for PICC-Related Bloodstream Infections in Premature Infants Using SHAP Interpretability

Objective To develop a risk prediction model for peripherally inserted central catheter-related bloodstream infections (PICC-CRBSI) in preterm infants. Methods We conducted a prospective study of 490 preterm infants with PICC placement in a Chinese tertiary neonatal ICU (June 2023–December 2024). The data were split into training and validation sets at a 7:3. CRBSI was the primary outcome. Feature selection was performed using LASSO regression, the Boruta algorithm, and recursive feature elimination (RFE), and prediction models were constructed using four machine learning algorithms, including logistic regression (LR), decision tree (DT), random forest (RF), and light gradient boosting machine (LightGBM). The optimal model was selected through performance comparison. Results CRBSI occurred in 68 cases (13.9%), predominantly caused by Gram-negative bacteria (e.g., Klebsiella pneumoniae, Escherichia coli). The random forest model performed best among the four machine learning models, with an area under the receiver operating characteristic curve (AUC) of 0.984, precision of 0.857, recall of 0.900, specificity of 0.976, accuracy of 0.966, and F1-score of 0.878. The SHAP summary plot identified the top 7 most important features in order: C-reactive protein (CRP), white blood cell count (WBC), respiratory rate, blood oxygen saturation, number of punctures, age at catheterization, and birth weight. Conclusion Our study developed and interpreted a risk prediction model for PICC- CRBSI in preterm infants. It is expected to assist clinicians in timely risk stratification and targeted interventions, thereby reducing the incidence of CRBSI and improving the prognosis of preterm infants. Clinical Trial Number: Not applicable.