Constructing a predictive model for early-onset sepsis in neonatal intensive care unit newborns based on SHapley...

Area of Science:

• Neonatal Medicine
• Computational Biology
• Medical Informatics

Background:

• Neonatal sepsis (NS) presents subtle, non-specific clinical signs, posing a significant threat to newborns.
• Early-onset sepsis (EOS), occurring within 72 hours of birth, is associated with high mortality rates.
• Accurate identification of NS risk factors and early diagnosis are critical for improving infant outcomes.

Purpose of the Study:

• Develop a robust machine learning (ML) model for early EOS prediction in neonatal intensive care units (NICUs).
• Identify pivotal risk factors contributing to EOS development.
• Provide interpretable insights into ML model predictions for clinical application.

Main Methods:

• Retrospective cohort study of 430 newborns (EOS and non-EOS) admitted to NICU.
• Data preprocessing and feature selection using LASSO regression.
• Evaluation of six ML models, including CatBoost, Random Forest, and XGBoost, using ROC and PR curves.
• Application of SHAP framework for model interpretability.

Main Results:

• All six ML models achieved ROCAUC > 0.900; CatBoost demonstrated superior performance (ROCAUC 0.975, PRAUC 0.947).
• Key identified risk factors for EOS include respiratory rate, procalcitonin, nasal congestion, yellow staining, white blood cell count, fever, and amniotic fluid turbidity.
• SHAP analysis provided ranked importance of these predictive features.

Conclusions:

• A precision-oriented ML model combined with SHAP interpretability effectively identifies critical EOS risk factors.
• This approach facilitates early prediction of EOS risk, enabling timely and targeted clinical interventions.
• The study supports the use of ML for precise diagnosis and treatment of neonatal sepsis.