Establishment of a Machine Learning-Based Prediction Model for Short-Term Adverse Prognosis in Neonatal Bacterial Meningitis

Background

Neonatal bacterial meningitis (NBM) is an extremely severe disease in the neonatal period. Early identification of high-risk infants is critical for timely intervention, yet prognostic assessment remains challenging due to nonspecific symptoms and variable clinical trajectories. While machine learning (ML) has shown promise in predicting outcomes for other neonatal conditions, its application for short-term adverse prognosis in NBM remains unexplored. This study aims to systematically evaluate ML models to screen for risk factors and identify the optimal predictive model to provide clinicians with a data-driven tool for the stratified management of NBM patients.

Methods

Clinical data of 433 term neonates with NBM hospitalized in the Department of Neonatology at the Capital Institute of Pediatrics between January 2013 and December 2023 were analyzed retrospectively. Based on discharge outcomes, patients were stratified into adverse (n = 84) and favorable prognosis (n = 349) groups. From an initial set of 32 clinical variables derived from clinical and laboratory data. Seventeen variables (15 via maximum Relevance Minimum Redundancy algorithm, two clinical-based) were selected. Nine machine learning models were evaluated using area under the curve (AUC), sensitivity, specificity, positive predictive value, and negative predictive value.

Results

Among the nine models, the logistic regression model achieved optimal performance (AUC: 0.908, accuracy: 0.890, sensitivity: 0.541, specificity: 0.974, positive predictive value: 0.845, negative predictive value: 0.898). Key predictors included muscle tone abnormalities, seizures, cerebrospinal fluid (CSF) protein > 2000 mg/L, mechanical ventilation, hypotension requiring inotropes, CSF glucose < 2.0 mmol/L, bulging fontanelle, C-reactive protein, hepatomegaly, and positive blood culture.

Conclusions

Machine learning models can be reliable tools for predicting short-term adverse prognoses in patients with NBM. The logistic regression model demonstrated the best predictive performance, which can help clinicians identify high-risk patients.