Development of a machine learning-based predictive model for long-term adverse outcomes in neonatal bacterial...

Area of Science:

• Neonatal medicine
• Computational biology
• Clinical informatics

Background:

• Neonatal bacterial meningitis (NBM) poses significant risks for long-term adverse prognosis.
• Early identification of high-risk infants is crucial for effective management and improved outcomes.
• Predictive models can aid clinicians in stratifying NBM patients based on prognosis.

Purpose of the Study:

• To apply machine learning for screening risk factors of long-term adverse prognosis in NBM.
• To develop and evaluate a robust prediction model for NBM outcomes.
• To identify key predictors for adverse prognosis in neonatal bacterial meningitis.

Main Methods:

• Included 139 neonates with NBM; divided into good (n=94) and poor (n=45) prognosis groups.
• Utilized Least Absolute Shrinkage and Selection Operator, Boruta, and Recursive Feature Elimination for feature selection.
• Constructed seven machine learning models, evaluating performance with AUC, accuracy, sensitivity, and specificity; interpreted using Shapley Additive explanation.

Main Results:

• The Random Forest model demonstrated superior clinical applicability with high accuracy (0.881), good calibration (Brier score: 0.123), and balanced sensitivity (0.887) and specificity (0.878).
• Logistic Regression showed high discriminative ability (AUC: 0.903).
• Key predictors identified include cerebrospinal fluid white blood cell count, cerebrospinal fluid protein levels, and seizures.

Conclusions:

• Machine learning models, especially Random Forest, reliably predict long-term adverse outcomes in NBM patients.
• These models assist in identifying infants at high risk for adverse prognosis.
• Further validation in diverse cohorts is recommended to improve generalizability and clinical use.