Machine learning models based on chest computed tomography for identifying plastic bronchitis in children with Mycoplasma pneumoniae pneumonia

Introduction

Mycoplasma pneumoniae (MP) is a prevalent pathogen of community-acquired pneumonia (CAP) in children, accounting for more than 40% of cases during epidemic years (1). Data from the World Health Organization (WHO) and many countries indicate a significant increase in children with MP pneumonia (MPP) after the global coronavirus disease 2019 (COVID-19) pandemic (2-5).

Plastic bronchitis (PB), as a term used to describe the presence of obstructing casts in the airways, can lead to life-threatening disorders if they occur in the large airways (6). PB is a major contributor to severe MPP and refractory MPP (RMPP) (7,8), and is the independent risk factor for obliterative bronchitis in children with RMPP (9). Fiberoptic bronchoscopy (FOB) and the bronchoalveolar lavage (BAL) were the standard for the diagnosis of PB in MPP. They have demonstrated significant efficacy in the diagnosis and treatment of PB, as they can remove bronchial casts (BCs) and alleviate airway obstruction (10,11). FOB should be performed as early as possible in case of severe MPP with PB (10). However, due to the lack of specific clinical features, the timing of PB formation remains unclear (12). In addition, as highly invasive procedures, FOB and BAL are not suitable for routine screening and have some relative contraindications, such as severe cardiopulmonary dysfunction, severe arrhythmia, high fever, and coagulation disorders (13). They are also associated with risks and complications, including hypoxia, bleeding, airway spasm, arrhythmia, and fever (14). Therefore, timely diagnosis of PB remains challenging yet indispensable for enhancing clinical outcomes.

Currently, chest computed tomography (CT) is one of the sensitive methods for diagnosing PB, as it can visualize the obstructing casts in the central airways with associated atelectasis and consolidation (15). However, the BCs may resemble other types of secretions, leading to misdiagnosis. In children with MPP, the BCs are usually small and located in subsegmental bronchi or invisible in lung consolidation or atelectasis areas, making them easy to overlook. For the clinical question of pneumonia, the chest CT guidelines were sparse and would typically recommend it in complicated pneumonia, poor treatment outcome or suspicion of complications. Usually, pediatricians can only determine the need for FOB based on the presence of consolidated or atelectatic areas of the lung (16). Machine learning (ML) based on chest CT may offer an effective approach for the early identification of PB.

ML can extract a large number of quantitative features from medical images, enabling the extraction of biological information about diseases by the analysis of these features (17). This approach combines medical imaging techniques with advanced data analysis methods, providing new perspectives and support for clinical decision-making (18). Recently, this approach has made remarkable progress in differentiating MPP, diagnosing RMPP and predicting macrolide resistance-associated gene mutation status of MPP (19-21). The results provide insight into the promising role of artificial intelligence in MPP. However, the utility of ML to identify PB has yet to be explored. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-545/rc).

Methods

Patients

This research was approved by the Institutional Review Board of Shandong Provincial Qianfoshan Hospital (No. 2025, S871). This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments, and the need for written informed consent was waived due to the retrospective nature of this study. The study cohort comprised eligible patients with complete clinical, CT imaging, laboratory, and FOB data, selected from three hospitals [The Second Qilu Hospital of Shandong University (Center 1), Shandong Provincial Hospital Affiliated to Shandong First Medical University (Center 2), and The First Affiliated Hospital of Shandong First Medical University (Center 3)] between January 2019 and October 2024, as shown in Table 1.

All patients were identified as having MPP based on laboratory tests (22). All had undergone both chest CT and FOB, and PB was diagnosed by FOB and BAL. The flowchart is shown in Figure 1.

Figure 1 The flowchart of patient recruitment. CT, computed tomography; FOB, fiberoptic bronchoscopy; MPP, Mycoplasma pneumoniae pneumonia; PB, plastic bronchitis.

The following conditions were required for inclusion:

• 28 days ≤ age ≤14 years.
• Positive laboratory results for MP infection, including:
  • Serological evidence: an MP-IgM titer of 1:160 or higher, or a four-fold increase in IgM titers between acute and convalescent serum samples;
  • Or molecular evidence: positive mediator probe polymerase chain reaction (MP-PCR) test results from nasopharyngeal swab, sputum, or BAL fluid;
• Confirmed the presence of pneumonia from the clinic and the image.
• FOB was needed and performed one or more times with image data and diagnostic reports.
• Chest CT imaging is performed within 5 days before the first FOB for complicated pneumonia, poor treatment outcome, or suspicion of complications.

Exclusion criteria were as follows:

• No FOB or CT images;
• No thin-slice CT images or poor-quality images;
• Absence of lung consolidation;
• Incomplete clinical or laboratory data;
• A disease course ≥4 weeks preadmission;
• Presence of underlying disease (e.g., congenital heart disease, bronchopulmonary dysplasia, congenital immunodeficiency disease) or chronic respiratory diseases (e.g., bronchiectasis, asthma).

CT image acquisition

The chest plain CT images we used were from different clinical CT systems, including SOMATOM Force CT, GE LightSpeed VCT, Optima 660, and Discovery CT750 HD. Imaging protocols were as follows: a tube voltage of 80–120 kV with automatic tube current modulation for dose optimization; depending on the scanner, the detector collimation of 192 mm × 0.6 mm or 128 mm × 0.6 mm with gantry rotation times between 0.25 and 0.4 seconds. Final images were reconstructed with a slice thickness and interval of either 1.0 or 1.25 mm. All CT images were reconstructed using a lung window (level: −500 HU, width: 1,400 HU) and a matrix size of 512×512.

Regions of interest (ROI) segmentation and feature extraction

To compare different image segmentation approaches, we defined the pneumonia-affected areas and the whole lung areas as ROI.

Digital Imaging and Communication in Medicine (DICOM) images were retrieved from the picture archiving and communication system, underwent anonymization and imported into the AI Research Portal software from Shenrui (https://www.radiomics.io/pyradiomics.html). The platform was employed to automatically generate the initial whole lung ROI and pneumonia ROI. An experienced radiologist then reviewed the automatically generated ROI and made manual adjustments as necessary. The final ROI assessment was independently reviewed by a senior radiologist to ensure agreement among readers. To address the limitations of manual ROI delineation, including high inter-individual variation and inefficiency, we utilized this semi-automated delineation approach. Figure 2 demonstrates a representative instance of the ROI segmentation process.

Figure 2 The workflow of the radiomics analysis. DCA, decision curve analysis; ICC, intraclass correlation coefficients; ROC, receiver operating characteristic; ROI, regions of interest.

Three-dimensional (3D) radiomics features were extracted from ROI of the entire image, using the AI Research Portal software from Shenrui. To reduce scanner and protocol-related heterogeneity and enhance the consistency and reproducibility of the extracted radiomics features, images underwent preprocessing before feature extraction. This included resampling to 1 mm³ voxels using trilinear interpolation and gray-level discretization with a bin width of 25 HU.

A total of 1,316 radiomics features were extracted in the whole lung ROI and the pneumonia ROI. These included first-order intensity histogram-based features, shape-based features, statistical matrix-based features, wavelet-based features and logarithmic transformations of 360 key metrics designed to enhance data distribution characteristics and mitigate the influence of outliers. After feature extraction, the Z-scores were used to normalize all radiomics data to facilitate subsequent analysis. The entire process is shown in Figure 2.

Additionally, to assess inter-observer reproducibility, a third radiologist independently re-contoured the region of interest in a randomly chosen subset of 40 patients. The stability of radiomic features was then evaluated using intraclass correlation coefficients (ICC). This step ensured that the selected features in the model were consistent across different raters, thereby enhancing the reliability of models.

Clinical data collection and model construction

Clinical medical data were collected at hospital admission, including demographic data (age, sex, fever, cough, peak body temperature and pleural effusion), laboratory tests [white blood cell (WBC), lymphocyte percentage (L%), neutrophil percentage (N%), neutrophil-lymphocyte ratio (NLR), C-reactive protein (CRP)]. Univariate and multivariable logistic regression analyses were performed to identify significant clinical risk factors, as shown in Table 2. Pleural effusion was the only variable significantly associated with PB and was therefore used to build the clinical prediction model (Model 1).

Radiomics features analysis and model construction

All three centers, affiliated with medical university, shared similar standards for disease diagnosis, treatment standards, clinical pathway, clinical practice indication, and operation specifications of CT and FOB. Accordingly, the children from Center 1 were assigned to the training set, Center 2 to the test set, and Center 3 to the validation set.

Radiomic feature selection was conducted through a multi-stage filtering process to ensure clinical relevance. First, only features demonstrating high reproducibility (ICC >0.75) were retained. Next, univariate statistical analysis using the two-sample t-test was applied to identify features significantly different between the PB and non-PB groups. Subsequently, least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation was employed to select the most predictive features, where the regularization parameter λ was optimized to minimize the mean squared error. Then, the rad-score of radiomics signature was constructed as a linear combination of the selected features, weighted by their respective LASSO coefficients: Rad-score = ∑ (coefficient i × feature i).

An optimal rad-score threshold was identified by maximizing the Youden index (specificity + sensitivity −1) to stratify patients into low- and high-risk radiomic categories. Finally, the whole lung model (Model 2) and the pneumonia model (Model 3) were established by logistic regression. The performance of these two models was then compared by DeLong test and Integrated Discrimination Improvement (IDI) index.

Multifactorial model construction and validation

To comprehensively assess patient’s condition, we combined the rad-score with clinical risk factors and performed univariate regression analysis to select significant features. We then built a multifactorial prediction model using multivariate logistic regression.

Finally, a total of four models were constructed to predict PB. As shown in Table 3 and Table 4, models were designated as follows: the clinical model based on the selected clinical factor (Model 1), the pneumonia CT model based on the pneumonia area radiomics labels (Model 2), the whole lung CT model based on the whole lung area radiomics labels (Model 3), and the multifactorial model which used the whole lung radiomics labels and selected clinical factor (Model 4). A flowchart of the model development is provided in Figure 2, outlining each stage from data preprocessing to final validation.

The receiver operating characteristic (ROC) curve and the area under the ROC curve were used to evaluate the predictive ability of the four models in the training, test, and validation sets. To compare the discriminatory performance between models, the DeLong test and IDI index were applied. Model calibration was assessed using calibration plots, while clinical usefulness was evaluated through decision curve analysis (DCA). Using the optimal model’s threshold, patients were stratified into low- and high-risk groups. In the training set, a nomogram was constructed to visualize the contribution of predictors and estimate individual patient risk, facilitating individualized risk prediction and therapeutic planning.

Statistical analysis

Continuous variables were summarized as means ± standard deviations. For three-group comparisons, the analysis of variance (ANOVA) or the Kruskal-Wallis test was used as appropriate, followed by Bonferroni-corrected post-hoc tests. For categorical variables, frequencies and percentages were used for description, and group comparisons were made using the Chi-squared (χ²) test or Fisher’s exact test, depending on the data characteristics. All statistical analyses were conducted using IBM SPSS Statistics (version 25), MedCalc (version 20.010), and R software (version 4.2.2). All statistical tests were two-sided, with statistical significance defined by a P value of <0.05.

Results

Study population and clinical characteristics

In Center 1, 3,945 children with MPP were excluded based on the exclusion criteria. A total of 373 children with MPP were included and assigned to the training set (mean age: 6.7±2.6 years; male/female: 188/185), and 137 patients (36.7%) were diagnosed as PB. In Center 2, 4,293 patients were excluded. A total of 221 patients (mean age: 6.1±2.8 years; male/female: 122/99) were included in the test set, in which 75 (33.9%) cases were PB. In Center 3, 4,320 patients were excluded. A total of 183 patients (mean age: 7.1±2.5 years; male/female: 89/94) were collected for the validation set, in which 68 (37.2%) cases were PB. Table 1 shows all clinical characteristics across the three sets. The patient inclusion workflow is illustrated in Figure 1.

Radiomics features analysis

Radiomic feature selection followed a standardized pipeline in both models. For the pneumonia model, 543 reproducible features (ICC >0.75) were identified and subsequently reduced to 505 via two-sample t-test (P<0.05). LASSO regression then selected 11 features, which formed the basis of the radiomics score (rad-score). In the whole lung model, 1,118 stable features (ICC >0.75) passed initial reliability screening and were further filtered to 328 significant features (P<0.05). The final rad-score was constructed using 7 features from wavelet and log-sigma filtered images identified by LASSO regression, with coefficients ranging from –0.023 to 0.087. These features capture texture and intensity variations associated with PB. The rad-score demonstrated strong discriminatory ability in predicting PB, as evidenced by ROC analysis in training set (Figure 3A), the test set (Figure 3B), and the validation set (Figure 3C). Among these features, the wavelet-LLL gray-level run-length matrix (GLRLM) Long Run High Gray Level Emphasis was identified as the most important predictor.

Figure 3 The ROC curves of Model 1 (clinical model), Model 2 (the pneumonia model), Model 3 (the whole lung model), and Model 4 (multifactorial model) in the training set (A), test set (B) and validation set (C). The multifactorial model obtained the highest AUC. AUC, area under the curve; CI, confidence interval; ROC, receiver operating characteristic.

Model building and testing

Univariable and multivariable logistic regression were employed to analyze the predictors of PB, with results shown in Table 2. Pleural effusion [odds ratio (OR), 3.268; 95% confidence interval (CI): 1.602–6.665, P=0.001], the pneumonia area radiomics labels (OR, 24.049; 95% CI: 1.017–568.599; P=0.049) and the whole lung area radiomics labels (OR, 54.893; 95% CI: 1.109–2,716.036; P=0.04) were independently associated with PB. These three risk factors were used to develop individual predictive models.

There was no significant difference in area under the curve (AUC) values between the whole lung model (Model 3) and the pneumonia model (Model 2) (P>0.05), as shown in Table 3. The whole lung ROI, including subtle radiological changes, enabled a thorough evaluation of the degree of lung lesions (23). For the good repeatability and conformability, the whole lung radiomics features were used for the multifactorial model in our further study. Table 4 shows the results for all models in the training, test, and validation sets. The ROC analyses, displayed in Figure 3A-3C, further illustrate their discriminative ability.

The multifactorial model achieved the best predicting performance for PB, which showed AUC values of 0.809 (95% CI: 0.763–0.855), 0.770 (95% CI: 0.702–0.839), and 0.831 (95% CI: 0.764–0.897) in the training set, the test set, and the validation set, respectively. Figure 4A-4C displayed the calibration of all models, demonstrating superior calibration performance for the multifactorial model. As shown in Figure 4D-4F, DCA demonstrated that this model yielded the greatest net benefit over a range of threshold probabilities.

Figure 4 The calibration curves and decision curves of four models. The calibration curves of four models in the training set (A), the test set (B) and the validation set (C). The decision curves for four models in the training set (D), the test set (E) and the validation set (F).

The sensitivity and specificity of the optimal PB prediction model were evaluated across three cohorts. In the training set, values reached 0.715 (95% CI: 0.632–0.789) and 0.767 (95% CI: 0.708–0.819), respectively. The test set yielded 0.720 (95% CI: 0.604–0.818) and 0.760 (95% CI: 0.683–0.827), while the validation set demonstrated 0.662 (95% CI: 0.537–0.772) and 0.878 (95% CI: 0.804–0.932). Optimal cut-off values for the rad-score were determined as 0.324 (the training set), 0.325 (the test set), and 0.342 (the validation set) using Youden’s index. A nomogram incorporating these predictors was developed based on the multifactorial model (Figure 5). Risk assessment via the nomogram indicates that the presence of pleural effusion and a high-risk radiomics label are associated with an increased probability of PB.

Figure 5 Nomogram of the multifactorial model.

Discussion

PB is a serious intrapulmonary complication of MPP (12). It is easy to be missed or misdiagnosed. Early assessment is important for optimizing treatment strategies and enhancing patient prognosis (6,11). In the present study, different prediction models were evolved and verified by ML. The results showed that the multifactorial model combined with radiological/clinical risk factors showed excellent performance in the training and validation sets, indicating its potential to early identify the MPP children at risk of developing PB. Furthermore, to compare different image segmentation approaches, we extracted radiomic features from both the regions of pneumonia and the whole lung. And we found that using the whole lung as an ROI might be the best image segmentation strategy for the study of pneumonia.

Our study aimed to develop an ML method based on chest CT to enable early identification of PB in children with MPP. Using this method, a large number of radiomics features and clinical indicators were automatically extracted and analyzed, thereby minimizing human selection bias and enhancing the objectivity and reproducibility of diagnostic assessment (18). To further ensure the generalizability of our study, the children’s data across three different hospitals were included. This multicenter data is better than the previous single-center nomogram studies, which lacked external data to validate their models (8,11).

In previous ML-based MPP studies, radiomics features were typically extracted from the pneumonia areas (19-21). According to CT findings, MPP can present with various imaging patterns, including bronchopneumonia, consolidation/atelectasis, bronchiolitis, and mosaic pattern (24). However, for bronchopneumonia, bronchiolitis, and mosaic patterns, the pattern often showed untidy margins, making it difficult to draw the outline. Moreover, some BCs located in small airways without obvious pneumonia may be overlooked. In contrast, the whole lung ROI can capture subtle radiological changes invisible to the naked eye and enable a thorough evaluation of the degree of lung lesions. This approach has been used in the previous studies (23,25). To compare the effectiveness of different image segmentation approaches, we defined both pneumonia areas and the whole lung as ROI. Due to poor repeatability and conformability in delineation, only 543 stable radiomics features in the pneumonia ROI were selected in ICC analysis, much fewer than the whole lung ROI in this study. Moreover, in the AUC values, our result showed no significant difference between the pneumonia model and the whole lung model based on chest CT. These findings suggest that using the whole-lung ROI may be a more robust image segmentation approach for studies of pneumonia, especially for beginners.

To evaluate the ability of chest CT in predicting PB in children with MPP complicated by lung consolidation, we extracted radiomics features from whole-lung CT images. Ultimately, seven radiomics features with non-zero coefficients were used to frame a whole-lung chest CT-associated predictive model. The results indicated that the model demonstrated good predictive performance in the training set (AUC =0.789), the test set (AUC =0.758), and the validation set (AUC =0.700), respectively. Among these features, the wavelet-LLL GLRLM Long Run High Gray Level Emphasis was identified as the most important predictor. This feature combines wavelet transformation with GLRLM analysis to capture detailed texture information from the CT images. It has been shown to be effective in distinguishing between different tissue types and disease states, such as identifying stages of pneumonia or monitoring treatment response in lung diseases (26). These findings support the feasibility of using CT image texture features to predict PB.

Early identification of PB remains challenging based on clinical features. Zhao et al. (8) and Zhang et al. (11) established different nomograms based on various clinical features and risk factors of PB caused by RMPP in children. They focused on some clinical characteristics and laboratory analysis data, which can be affected by factors such as the infection time, prior use of antibiotic drugs before hospitalization, and transfer treatment. These variabilities make it hard to ensure consistency in different children. Moreover, both studies only focused on PB with RMPP. In contrast, our present study minimized inconsistencies in clinical characteristics and laboratory data, and pleural effusion was the only clinical indicator selected to build the clinical model. Pleural effusion was often associated with the severity of pneumonia, and was recognized as a risk factor for PB (11,27,28). In MP infection, the inflammatory response triggers the release of mediators, that not only increase the permeability of blood vessels but also contribute to the thickening of the pleural lining. Both of them were key factors in the development of pleural effusion (27). Pleural effusion can cause lung compression and restrictive ventilation dysfunction, which may worsen respiratory function in PB patients (11). Thus, pleural effusion must be closely monitored and carefully managed in pneumonia patients (28).

Regarding model comparison, the multifactorial model combining clinical information and whole-lung radiomics features exhibited the best overall performance for identification of PB. This demonstrates that the fusion of multi-source information contributes to enhancing diagnostic accuracy and reliability. Through the non-invasive method, potential PB children could be identified early. On the one hand, the prompt utilization of FOB and BAL for PB children can remove BCs, enhance lung ventilation and eliminate various inflammatory factors, thus accelerating clinical symptoms resolution and shortening hospitalization duration (10). On the other hand, the FOB for non-PB children with lung consolidation could be avoided for the surgical trauma and postoperative complications (13,14). If the calculated risk of PB formation in a patient was low, the clinicians may choose to monitor. However, if the risk assessment indicated a high likelihood of PB, more aggressive treatment strategies and FOB may be warranted (10). Therefore, the multiple models, combining radiomics with clinical indicators, may offer new tools for the early diagnosis of MPP-related complications.

While this work provides valuable insights, its retrospective design, limited sample size, and imaging protocol heterogeneity across participating centers represent notable limitations. The exclusion of patients with incomplete data may have skewed the cohort, and differences in CT acquisition parameters could affect the robustness of radiomic features. Prospective studies with larger, more homogeneous datasets and standardized imaging protocols are needed to validate and extend these findings.

Conclusions

A non-invasive ML multifactorial model was evolved by combining chest CT radiomics features with clinical risk factors, and it demonstrated favorable accuracy for the early identification of PB in children with MPP prior to FOB. This model may contribute to timely intervention and appropriate treatment.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-545/rc

Data Sharing Statement: Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-545/dss

Peer Review File: Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-545/prf

Funding: This study was supported by Shandong Provincial Medical and Health Scientific and Technological Project (grant No. 2019WS509).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-545/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This research was approved by the Institutional Review Board of Shandong Provincial Qianfoshan Hospital (No. 2025, S871). This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments, and the need for written informed consent was waived due to the retrospective nature of this study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Song Z, Jia G, Luo G, et al. Global research trends of Mycoplasma pneumoniae pneumonia in children: a bibliometric analysis. Front Pediatr 2023;11:1306234. [Crossref] [PubMed]
2. Meyer Sauteur PM, Beeton MLEuropean Society of Clinical Microbiology and Infectious Diseases (ESCMID) Study Group for Mycoplasma and Chlamydia Infections (ESGMAC), et al. Pneumonia outbreaks due to re-emergence of Mycoplasma pneumoniae. Lancet Microbe 2024;5:e514. [Crossref] [PubMed]
3. Edens C, Clopper BR, DeVies J, et al. Notes from the Field: Reemergence of Mycoplasma pneumoniae Infections in Children and Adolescents After the COVID-19 Pandemic, United States, 2018-2024. MMWR Morb Mortal Wkly Rep 2024;73:149-51. [Crossref] [PubMed]
4. Liang Y, Zhang P, Du B, et al. Changes of Mycoplasma pneumoniae infection in children before and after the COVID - 19 pandemic, Henan, China. J Infect 2023;86:256-308. [Crossref] [PubMed]
5. Meyer Sauteur PM, Beeton MLEuropean Society of Clinical Microbiology and Infectious Diseases (ESCMID) Study Group for Mycoplasma and Chlamydia Infections (ESGMAC). Mycoplasma pneumoniae: delayed re-emergence after COVID-19 pandemic restrictions. Lancet Microbe 2024;5:e100-1. [Crossref] [PubMed]
6. Yang L, Zhang Y, Shen C, et al. Clinical features and risk factors of plastic bronchitis caused by Mycoplasma pneumoniae pneumonia in children. BMC Pulm Med 2023;23:468. [Crossref] [PubMed]
7. Mu SY, Zou YX, Guo YS, et al. Clinical characteristics and predictive factors for plastic bronchitis in children with severe Mycoplasma pneumoniae pneumonia. Zhonghua Er Ke Za Zhi 2024;62:861-6. [Crossref] [PubMed]
8. Zhao L, Zhang T, Cui X, et al. Development and validation of a nomogram to predict plastic bronchitis in children with refractory Mycoplasma pneumoniae pneumonia. BMC Pulm Med 2022;22:253. [Crossref] [PubMed]
9. Cheng Q, Zhang H, Shang Y, et al. Clinical features and risk factors analysis of bronchitis obliterans due to refractory Mycoplasma pneumoniae pneumonia in children: a nomogram prediction model. BMC Infect Dis 2021;21:1085. [Crossref] [PubMed]
10. Wang L, Wang W, Sun JM, et al. Efficacy of fiberoptic bronchoscopy and bronchoalveolar lavage in childhood-onset, complicated plastic bronchitis. Pediatr Pulmonol 2020;55:3088-95. [Crossref] [PubMed]
11. Zhang H, Yang J, Zhao W, et al. Clinical features and risk factors of plastic bronchitis caused by refractory Mycoplasma pneumoniae pneumonia in children: a practical nomogram prediction model. Eur J Pediatr 2023;182:1239-49. [Crossref] [PubMed]
12. Huang JJ, Yang XQ, Zhuo ZQ, et al. Clinical characteristics of plastic bronchitis in children: a retrospective analysis of 43 cases. Respir Res 2022;23:51. [Crossref] [PubMed]
13. Yavuz S, Sherif A, Saif S, et al. Indications, Efficacy, and Complications of Pediatric Bronchoscopy: A Retrospective Study at a Tertiary Center. Cureus 2023;15:e40888. [Crossref] [PubMed]
14. Midulla F, de Blic J, Barbato A, et al. Flexible endoscopy of paediatric airways. Eur Respir J 2003;22:698-708. [Crossref] [PubMed]
15. Liu F, Zhao L, Gao Y, et al. Case Report: Aspergillus fumigatus-associated plastic bronchitis in a pediatric patient with cystic fibrosis complicated by allergic bronchopulmonary aspergillosis. Front Pediatr 2025;13:1632989. [Crossref] [PubMed]
16. Lu W, Wu X, Xu Y, et al. Predictive value of bronchoscopy combined with CT score for refractory mycoplasma pneumoniae pneumonia in children. BMC Pulm Med 2024;24:251. [Crossref] [PubMed]
17. Greener JG, Kandathil SM, Moffat L, et al. A guide to machine learning for biologists. Nat Rev Mol Cell Biol 2022;23:40-55. [Crossref] [PubMed]
18. Handelman GS, Kok HK, Chandra RV, et al. eDoctor: machine learning and the future of medicine. J Intern Med 2018;284:603-19. [Crossref] [PubMed]
19. Qian Y, Tao Y, Wu L, et al. Model based on the automated AI-driven CT quantification is effective for the diagnosis of refractory Mycoplasma pneumoniae pneumonia. Sci Rep 2024;14:16172. [Crossref] [PubMed]
20. Wang D, Zhao J, Zhang R, et al. The value of CT radiomic in differentiating mycoplasma pneumoniae pneumonia from streptococcus pneumoniae pneumonia with similar consolidation in children under 5 years. Front Pediatr 2022;10:953399. [Crossref] [PubMed]
21. Wang H, Zhao L, Yang Y, et al. Chest Computed Tomography Radiomics for Determining Macrolide Resistance-Associated Gene Mutation Status in Children with Mycoplasma pneumoniae Pneumonia: A Two-Center Study. Acad Radiol 2024;31:3774-82. [Crossref] [PubMed]
22. Gao L, Sun Y. Laboratory diagnosis and treatment of Mycoplasma pneumoniae infection in children: a review. Ann Med 2024;56:2386636. [Crossref] [PubMed]
23. Yanling W, Duo G, Zuojun G, et al. Radiomics Nomogram Analyses for Differentiating Pneumonia and Acute Paraquat Lung Injury. Sci Rep 2019;9:15029. [Crossref] [PubMed]
24. Huang X, Gu H, Wu R, et al. Chest imaging classification in Mycoplasma pneumoniae pneumonia is associated with its clinical features and outcomes. Respir Med 2024;221:107480. [Crossref] [PubMed]
25. Liu J, Zhang J, Wang H, et al. Machine Learning Methods Based on Chest CT for Predicting the Risk of COVID-19-Associated Pulmonary Aspergillosis. Acad Radiol 2025;32:3725-38. [Crossref] [PubMed]
26. Li L, Wang L, Zeng F, et al. Development and multicenter validation of a CT-based radiomics signature for predicting severe COVID-19 pneumonia. Eur Radiol 2021;31:7901-12. [Crossref] [PubMed]
27. Ji L, Lin L, Ye J, et al. Clinical Characteristics and Predictive Factors Analysis of Mycoplasma Pneumoniae Pneumonia Complicated with Pleural Effusion in Children. Iran J Allergy Asthma Immunol 2025;24:451-61.
28. Lin X, Xu E, Zhang T, et al. Risk factors for pleural effusion in children with plastic bronchitis caused by pneumonia. Front Pediatr 2024;12:1412729. [Crossref] [PubMed]