Abstract:
Objectives: Neonatal bacterial meningitis (NBM) is a severe invasive infection of the central nervous system and a leading cause of early-life mortality and long-term neurodevelopmental impairment. Although its case-fatality ratio has declined over time, the burden of mortality and neurologic morbidity remains substantial. Survivors often face persistent developmental deficits and increased healthcare utilization. This study aimed to develop and externally validate a clinical prediction model based on routinely available clinical indicators for early identification of term neonates with bacterial meningitis at high risk of imaging-defined intracranial complications, thereby enabling targeted neurological surveillance and timely imaging follow-up.
Methods: This retrospective two-center study included term neonates with bacterial meningitis who remained hospitalized for at least 14 days. Thirty prespecified predictors available within the first three days after symptom onset were evaluated. Predictor stability was assessed within a leakage-controlled resampling framework, six algorithms were compared using identical validation partitions, and the external cohort remained isolated until the development procedure and final model were frozen.
Results: The development cohort included 565 infants, of whom 138 (24.4%) developed imaging-defined complications; the external cohort included 102 infants, with 28 (27.5%) events. The final model included cerebrospinal fluid white blood cell count, altered muscle tone, hypotension requiring vasoactive support, gaze deviation, seizure, and C-reactive protein. Nested validation of the complete development procedure yielded an AUROC of 0.806, a Brier score of 0.137, and a calibration slope of 0.838. Conditional on the final predictor set, logistic regression had the lowest log loss and Brier score, and no complex algorithm provided a consistent overall advantage. External validation yielded an AUROC of 0.637, a Brier score of 0.233, and a calibration slope of 0.286.
Conclusions: The six-predictor logistic regression model summarized clinically coherent early prognostic information, but external discrimination was modest and calibration was poor. It should be considered a candidate framework for multicenter validation and independently evaluated updating rather than a model ready for individual clinical decision-making.
