A Web-Based Dynamic Nomogram for Predicting Refractory Mycoplasma pneumoniae Pneumonia in Children

Summary

The present study comprehensively evaluated clinical manifestations, laboratory indices, and imaging features in children with MPP to identify independent risk factors for RMPP and to establish an online dynamic nomogram for early prediction.

Abstract

Refractory Mycoplasma pneumoniae pneumonia (RMPP) in children is associated with stronger inflammatory responses, more complex management, and a higher risk of pulmonary complications. Early identification of high-risk children may support timely treatment adjustment. This retrospective study included 500 hospitalized children with Mycoplasma pneumoniae pneumonia from November 2022 to November 2023. The cohort was randomly divided into a training set (n = 375) and a validation set (n = 125). Univariable and multivariable logistic regression analyses identified six independent predictors of RMPP: fever duration before admission, peak body temperature, lactate dehydrogenase level, sputum plugs or pleural effusion, lung consolidation, and hypoxemia. A web-based dynamic nomogram was constructed using these variables. The model showed good discrimination, with an area under the receiver operating characteristic curve/C-index of 0.840 (95% CI, 0.785–0.892) in the training set and 0.831 (95% CI, 0.743–0.906) in the validation set. Calibration curves showed good agreement between predicted and observed risks, and decision curve analysis suggested clinical net benefit across most threshold probabilities. Children with prolonged fever, high peak temperature, elevated lactate dehydrogenase, hypoxemia, and CT evidence of lung consolidation or sputum plugs/pleural effusion should be considered at increased risk of RMPP and may require closer monitoring and timely treatment adjustment.

Introduction

Mycoplasma pneumoniae (MP) is a major pathogen of community-acquired pneumonia in children, especially in preschool-and school-aged populations¹. Mycoplasma pneumoniae pneumonia (MPP) accounts for a considerable proportion of pediatric pneumonia, and its incidence has increased in recent years². Most children have a favorable course after appropriate treatment; however, some develop persistent fever, worsening respiratory symptoms, or progressive radiographic abnormalities after at least 7 days of standard macrolide therapy, a phenomenon potentially attributable to macrolide resistance and/or delayed treatment. This condition is generally defined as refractory Mycoplasma pneumoniae pneumonia (RMPP). RMPP is associated with faster progression, more complex treatment, and a higher risk of pulmonary and extrapulmonary complications, making early risk identification clinically important.

Several clinical, laboratory, and imaging indicators have been associated with RMPP. Reported predictors include lactate dehydrogenase (LDH), D-dimer, interleukin (IL)-6, IL-10, pleural effusion, mucus plugging, pulmonary consolidation, prolonged fever, high peak temperature, older age, and hypoxemia^{3,4,5,6,7,8,9}. These variables reflect different aspects of disease activity, including inflammatory response, tissue injury, coagulation activation, airway obstruction, and impaired oxygenation. However, RMPP is usually driven by multiple interacting factors, and single-variable assessment may provide limited individualized risk estimation.

Current clinical prediction mainly relies on physician judgment, threshold-based laboratory interpretation, imaging assessment, or static scoring models. Clinical judgment is flexible but subjective, whereas single-indicator thresholds are simple but may miss the combined effect of clinical severity, inflammation, and radiographic progression. A nomogram integrates multiple independent predictors into an individualized probability estimate and provides a more intuitive risk assessment than isolated variables. A web-based dynamic nomogram further allows direct input of patient data and rapid bedside calculation, which may support early monitoring and treatment adjustment.

Prediction models remain decision-support tools and cannot replace clinical judgment. Their performance depends on data quality, variable definitions, and validation in independent populations. Existing nomogram-based studies for RMPP have been reported^6,7; however, externally validated dynamic tools and systematic comparisons across models are still needed. The web-based nomogram developed in this study is intended for use in pediatric departments and emergency settings where chest CT, LDH measurement, and oxygen saturation monitoring are routinely available. Clinicians should interpret predictions with caution when applied to populations with markedly different epidemiological profiles, healthcare settings with limited imaging resources, or patients receiving pre-admission corticosteroids or second-line antibiotics, which may alter inflammatory markers.

Protocol

This study was approved by the Ethics Committee of Jinhua Women's and Children's Hospital (No. 2024KY099). Informed consent was obtained from the children's legal guardians.

Patient data

This retrospective study included children with Mycoplasma pneumoniae pneumonia (MPP) hospitalized at the Children’s Hospital Affiliated to Jinhua Women and Children’s Hospital from November 2022 to November 2023. MPP was diagnosed based on respiratory symptoms, chest imaging findings, and positive Mycoplasma pneumoniae DNA or RNA testing. General MPP (GMPP) refers to children who responded adequately to standard macrolide therapy, whereas RMPP was defined as persistent fever, aggravated symptoms, or progressive imaging abnormalities after at least 7 days of standard macrolide therapy¹⁰. The inclusion criteria were age <14 years, confirmed Mycoplasma pneumoniae infection, respiratory symptoms, and imaging findings consistent with pneumonia. The exclusion criteria were mixed infection within 10 days after disease onset, leukemia, chronic lung disease, immunodeficiency, previous immunosuppressive therapy, admission during the recovery phase, or incomplete records. Standard therapy referred to azithromycin 10 mg/kg once daily, orally or intravenously, with a maximum dose of 500 mg/day. Children not meeting the RMPP criteria were classified as general MPP (GMPP). A total of 500 eligible children were randomly divided into a training cohort (n = 375) and a validation cohort (n = 125) using R software version 4.1.2 (function: sample(), set.seed = 42). The validation cohort was reserved exclusively for external model evaluation and was not used in any step of model construction. Electronic medical record data were extracted using a standardized extraction template developed a priori. Data fields included demographics, clinical presentation, laboratory values, imaging findings, and treatment records. Two trained investigators independently extracted the data, and discrepancies were resolved by consensus. Chest CT findings were assessed by radiologists blinded to clinical grouping and laboratory results. Inter-rater reliability for CT image interpretation was assessed using Cohen's kappa statistic for binary imaging findings (sputum plugs or pleural effusion: yes/no; lung consolidation: yes/no). Finally, 129 children were classified as RMPP and 371 as general MPP.

Variables

Demographic, clinical, laboratory, and imaging variables were extracted from electronic medical records using a predefined form. Clinical variables included age, sex, fever duration before admission, peak body temperature, and hypoxemia. Fever duration was defined as the interval from fever onset to admission. Peak body temperature was the highest recorded temperature before or within 24 h after admission. Hypoxemia was defined as peripheral arterial oxygen saturation <92% on room air or the need for supplemental oxygen. Laboratory variables included WBC, HB, PLT, CRP, ALB, ALT, CK-MB, LDH, D-dimer, IL-6, IL-8, IL-10, IL-17, PCT, and NE%. Fasting venous blood samples were collected within 24 h after admission, and sampling time relative to fever onset and antibiotic initiation was recorded when available. Mycoplasma pneumoniae DNA or RNA testing was used for etiological confirmation. The throat swab real-time quantitative PCR assay for Mycoplasma pneumoniae DNA or RNA was positive. Quantitative DNA/RNA load was not included because standardized load data were unavailable for all patients. Chest CT performed within 3 days before or after admission was reviewed for sputum plugs, pleural effusion, and lung consolidation. “Sputum plugs or pleural effusion” was recorded as positive when either finding was present. Lung consolidation was defined as segmental or lobar parenchymal opacity on CT. A 256-slice CT scanner was used. Children were placed in the supine position, and low-dose axial chest scanning was performed with a detector of appropriate width under breath-holding or while the children were asleep. The scan coverage extended from the lung apex to the lung base to cover the entire lung parenchyma. Scan parameters were set as follows: tube voltage 100 kVp, slice thickness 5 mm, reconstruction slice thickness 1.25 mm, matrix 512 × 512, gantry rotation time 0.28 s, noise index 12. Images were reconstructed using ASIR-V and DLIR algorithms. Each group contained reconstructed images with ASIR-V weights of 20%, 50%, and 80%, as well as DLIR-L, DLIR-M, and DLIR-H images. The acquired image data were imported into the workstation. All images were independently reviewed in a double-blind manner by two radiologists with more than five years of clinical experience. They assessed the presence of abnormal pulmonary signs and summarized the main imaging features. For cases with discrepant interpretations, a consensus was reached through mutual discussion.

Statistical analysis

Statistical analyses were performed using R software version 4.1.2. Categorical variables are presented as n (%) and compared using the chi-square test or Fisher’s exact test. Continuous variables are expressed as mean ± standard deviation or median (interquartile range), according to distribution. Normality was assessed using the Shapiro-Wilk test. Normally distributed variables were compared using the independent-samples t test, and non-normally distributed variables using the Mann-Whitney U test. Before model construction, baseline comparability between the training and validation cohorts was confirmed using χ² or Mann-Whitney U tests for all candidate predictors, demonstrating no significant between-cohort differences (all P > 0.05). In the training cohort, variables with P < 0.05 in univariate logistic regression were entered into multivariable logistic regression. Multicollinearity was assessed using variance inflation factors (VIFs) calculated with the "car" package; all selected predictors had VIF < 5, indicating no substantial multicollinearity. The linearity assumption for LDH was examined using logarithmic transformation and restricted cubic splines (using the "rms" package with 3 knots placed at the 10th, 50th, and 90th percentiles); the linear form was retained when model fit did not improve (likelihood ratio test: P > 0.05 for the non-linear spline term, supporting the linear specification). A nomogram was constructed from the final multivariable model using the "rms" package (version 6.3-0; functions: lrm for logistic regression, Predict for value prediction, and nomogram for graphical representation). Discrimination was evaluated using receiver operating characteristic curves, area under the curve, and C-index with 95% confidence intervals ("pROC" package). Calibration was assessed using calibration curves with 1,000 bootstrap resamples. ("rms" package: calibrate function). The optimal prediction cutoff was determined independently within each cohort using Youden's index via the pROC package: 0.222 in the training cohort and 0.247 in the validation cohort. The training-derived cutoff (0.222) was also applied to the validation cohort for cross-cohort performance comparison, consistent with standard internal validation practice. The 95% confidence interval for the predicted probability, for example, was calculated using model-based standard errors on the logit scale. Decision curve analysis was used to evaluate net benefit across threshold probabilities ("rmda" package: decision_curve and plot_decision_curve functions). A two-sided P < 0.05 was considered statistically significant. The complete R scripts, including data preprocessing, model construction, validation, and nomogram generation, are available from the corresponding author upon request.

Missing data were minimal (all variables <2% missing) and handled using median imputation for continuous variables and mode imputation for categorical variables prior to analysis. All categorical predictors were coded as binary (0/1) indicator variables. Continuous predictors (fever duration, peak temperature, LDH) were retained in their original clinical units without categorization. No variable selection based on univariate screening was performed prior to multivariable modeling; instead, all clinically relevant candidates identified in the literature were considered, and those with P < 0.05 in the univariate analysis of the training cohort were advanced to the multivariable model.

Web calculator implementation

The web-based dynamic nomogram was deployed using the "shiny" package (version 1.7.4) in R. The user interface was built with shiny::fluidPage(), shiny::sidebarLayout(), and shiny::sliderInput() for continuous variables (fever duration, peak temperature, LDH) and shiny::selectInput() for binary variables (sputum plugs or pleural effusion, lung consolidation, hypoxemia). The server logic invoked the predict() function from the fitted lrm model to compute individual risk estimates, with the resulting probabilities rendered via renderPlot() and renderText(). The application was hosted on ShinyApps.io (https://predictrmpp.shinyapps.io/RMPP/).

Results

Clinical data were collected retrospectively and handled in accordance with institutional confidentiality requirements.

Study population

A total of 500 eligible children were included, including 272 boys and 228 girls. The training cohort included 375 children, of whom 286 had GMPP (general MPP), and 89 had RMPP. The validation cohort included 125 children, of whom 85 had GMPP, and 40 had RMPP. The training and validation cohorts were comparable in ...

Discussion

RMPP is associated with sustained inflammation, poor response to standard macrolide therapy, and increased risk of pulmonary and extrapulmonary complications. Early identification of high-risk children is therefore important for timely treatment adjustment. In this study, fever duration before admission, peak body temperature, LDH, sputum plugs or pleural effusion, lung consolidation, and hypoxemia were independently associated with RMPP. A web-based dynamic nomogram based on these variables showed stable discrimination,...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
R software version 4.1.2	R Foundation for Statistical Computing, Vienna, Austria	http://www.r-project.org
Revolution 256-slice CT scanner	GE Healthcare, United States	www.gehealthcare.com