Machine learning models for predicting neonatal bacterial infections: a retrospective cohort study

Area of Science:

• Neonatal Medicine
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Bacterial infections are a major cause of neonatal mortality globally.
• Early and accurate diagnosis in infants (1-90 days) is crucial for timely intervention.
• Current diagnostic methods can be invasive, necessitating less invasive alternatives.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for early, non-invasive prediction of bacterial infections in neonates.
• To utilize routine clinical data for point-of-care triage and risk stratification.
• To ensure clinical interpretability of the developed ML models.

Main Methods:

• Retrospective analysis of clinical data from 306 infants (1-90 days old).
• Target variable defined by cerebrospinal fluid (CSF) culture results.
• Nine ML classifiers evaluated using a nested cross-validation (NCV) framework and SHapley Additive exPlanations (SHAP) for interpretability.

Main Results:

• Top ML models achieved an AUROC of 0.74-0.76.
• L2-regularized logistic regression (LR) was selected for its stable discrimination and minimal generalization gap (0.083).
• SHAP analysis indicated predictive value from urinary markers, age, and metabolic indicators.

Conclusions:

• ML models using routine, non-invasive paraclinical markers can assist in risk-stratifying infants for bacterial infections.
• Threshold-dependent specificity trade-offs necessitate using these models as clinical decision support tools, not standalone rule-out methods.
• The optimized pipeline aids risk tiering, complementing rather than replacing invasive diagnostic procedures like lumbar punctures.