To analyze the clinical characteristics of children with invasive pneumococcal disease (IPD) admitted to Kunming Children’s Hospital from 2020 to 2024, and to develop a risk stratification model for pneumococcal meningitis among children with confirmed or highly suspected IPD using multiple machine learning algorithms. A retrospective analysis was conducted on 65 pediatric patients with confirmed IPD, divided into a meningitis group (n = 37) and a non-meningitis group (n = 28). Clinical characteristics between the two groups were compared. Variable selection was performed using LASSO regression, and six machine learning models—Logistic Regression (LR), K-Nearest Neighbor (KNN), Naive Bayes (NB), Multilayer Perceptron (MLP), Random Forest (RF), and XGBoost—were constructed based on the selected features. Model performance was evaluated using AUC, F1 score, accuracy, sensitivity, and specificity, while Decision Curve Analysis (DCA) was employed to assess clinical utility. The meningitis group exhibited stronger inflammatory responses and poorer outcomes (in-hospital death and treatment abandonment). The predominant serotypes were 19 F (25.8%), 19 A (20.9%), and 14 (17.7%). LASSO regression identified six key predictive variables: headache, vomiting, nuchal rigidity, disease course, C-reactive protein (CRP), and blood urea nitrogen (BUN). Among the machine learning models, Logistic Regression and MLP performed best, with AUC values of 0.942 and 0.947, respectively. DCA indicated the highest net clinical benefit for these two models. A nomogram based on the Logistic Regression model enabled individualized risk estimation within the IPD population. Pediatric pneumococcal meningitis remains associated with high disability rates and poor prognosis. The machine learning-based predictive model integrating clinical symptoms and laboratory indicators demonstrated a discriminative ability and may hold the potential to serve as a supportive tool for risk stratification among children with invasive pneumococcal disease. Not applicable.
Luo et al. (Tue,) studied this question.