The development and validation of a privacy-preserving model based on federated learning for diagnosing severe pediatric pneumonia

Introduction

Pediatric pneumonia is the most common respiratory disease in children (1-3). Most infants and young children experience pneumonia during the winter and spring seasons, with 7% to 13% of cases being severe pneumonia with a rapid progression and obvious symptoms of systemic poisoning. In addition to respiratory symptoms, acute gastrointestinal dysfunction, acute respiratory failure, microcirculation failure, myocarditis, respiratory insufficiency, moderate toxic encephalopathy, sepsis, and water electrolyte imbalance can also occur (4).

Insufficient understanding of the disease, lack of rational treatment in the early stages, and the critical condition at the time of treatment are the high-risk factors for severe pneumonia. The treatment of severe pneumonia includes regular oxygen therapy, mechanical ventilation, antibiotic therapy, and symptomatic treatment (4,5). If not treated in a timely manner, pneumonia can lead to death.

At present, there is a challenge of in diagnostic testing of pneumonia in children, especially severe pneumonia (6). Thus, developing an auxiliary diagnostic model to help identify severe pneumonia in pediatric patients at an early stage would be highly valuable. In recent years, several diagnostic models for pediatric pneumonia have been proposed. Chang et al. (7) developed a model to predict severe Mycoplasma pneumoniae pneumonia in pediatric patients. Haggie et al. (8) established a model to identify severe disease among patients with pediatric pneumonia. Li et al. (9) developed a predictive nomogram model for refractory Mycoplasma pneumoniae pneumonia in hospitalized children. Katreddi et al. (10) developed a model to predict pediatric pneumonia based on X-ray images. Although these models demonstrated good discrimination and calibration, providing a basis for the early identification of severe pediatric pneumonia, their efficacy should be validated by multicenter studies in a large sample.

Multicenter studies have several advantages over single-center studies, including a sufficient data size and the improved generalizability and reproducibility of the research outcomes (11-13). However, many medical centers are unwilling to share their data due to heightened data security risks, which hinders the collaborative benefits of multicenter research (14). In this study, we employed the Arya privacy computing platform (Hangzhou Healink Technology Corporation, Hangzhou, China) to conduct a multicenter retrospective study and developed a privacy-preserving federated learning model to identify severe pneumonia in pediatric patients. The diagnostic model can benefit doctors and patients as an auxiliary diagnostic tool. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-349/rc).

Methods

Workflow of federated learning on the Ayra platform

The workflow of federated learning on the Arya platform is summarized in Figure 1. In this process, Arya fist helps to build a single data platform with each hospital as the core. The patient examination data are processed into the biochemical indicators’ dataset (the data cannot be shared due to privacy issues). Subsequently, a data collaboration alliance network is built, which provides two-way promotion. Each hospital client can upload a local model to the central server in each round, which aggregates the global model and returns it to each client. Each hospital client trains the local data on the basis of the global model, which is an iteration process. Finally, the global pneumonia detection model is obtained after a specified number of iterations.

Figure 1 The workflow of federated learning on the Ayra platform.

Datasets

This multicenter retrospective study was approved by the Medical Ethics Committee of Children’s Hospital, Zhejiang University School of Medicine (No. 2023-IRB-0181-P-01) and the requirement for written informed consent from patients was waived as long as the data of the patients’ remained anonymous. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. All participating hospitals/institutions were informed and agreed with the study.

Patients were included in the study based on their discharge International Classification of Diseases, 10^th Revision (ICD-10) codes recorded in the hospital information system. Severe pneumonia cases were identified using the ICD-10 code J18.903, which denotes pneumonia of unspecified organism with severe clinical presentation, according to the Chinese Clinical Modification of the ICD-10. Common pneumonia cases were defined by the codes J18.000a and J18.900a.

The dataset was derived from four medical centers, including Children’s Hospital, Zhejiang University School of Medicine, Children’s Hospital of Soochow University, Changchun Children’s Hospital, and Kunming Children’s Hospital. A total of 5,091 records (each record corresponding to a pediatric patient diagnosed with pneumonia) were included, with 2,484 cases of severe pneumonia and 2,607 with common pneumonia. The details of the datasets from the four hospitals are listed in Table 1. The data were not merged; but stored in the information center of each hospital separately . A total of 33 biochemical and immune indicators were collected: white blood cell count (WBC), high-sensitivity C-reactive protein (hs-CRP), hemoglobin (Hb), platelet count (PLT), lymphocyte percentage (L%), monocyte percentage (M%), neutrophil percentage (N%), erythrocyte sedimentation rate (ESR), procalcitonin (PCT), D-dimer, prothrombin time (PT), activated partial thromboplastin time (APTT),alanine aminotransferase (ALT), aspartate aminotransferase (AST), lactic dehydrogenase (LDH), interferon-r (IFN-r), interleukin-2 (IL-2), interleukin-4 (IL-4), interleukin (IL-6), interleukin (IL-10), tumor necrosis factor (TNF), and T-cell subpopulations (CD25/CD3⁺CD8⁺, CD25⁺/CD4⁺CD3⁺, CD19, CD3, CD3-CD16⁺CD56⁺, CD4, CD4/CD8, CD69⁺/CD3⁺CD4⁺, CD69⁺/CD3⁺CD8⁺, CD8, HLA-DR⁺/CD3⁺CD4⁺, HLA-DR⁺/CD3⁺CD8⁺). As shown in Figure 2, the dataset was divided into a training set and test set in a hospital-level stratified manner. Among the records, 80% were randomly selected for model training, and the remaining 20% records were used in the testing phase.

Figure 2 The details of data splitting.

Identification of common features

For data training, it was necessary to screen common features (indicators) from the data collected by different hospitals. The unique hospital indicators were removed to ensure data consistency and comparability. After screening, a total of 11 indicators were ultimately identified for training: WBC, hs-CRP, Hb, PLT, L%, M%, N%, PT, ALT, AST, and LDH.

Data discretization processing

Due to the different instruments used for the detection of biochemical indicators, the original values could not be directly used for model training, as it could mislead the model and lead to poor results. Therefore, data discretization processing was performed to convert continuous values into finite class values, which could simplify the complexity of data, reduce data noise, and enhance the interpretability and visualization of data. The details of this method are as follows: a value below the normal range is treated as −1, indicating that the indicator is low; a value within the normal range is treated as 0, indicating that the indicator is normal; and a value over the normal range is treated as 1, indicating that the indicator is high. Through this processing method, we could convert the value of each indicator into a ternary symbol (−1, 0, 1).

Data filling

Not all patients had complete data due to different queries from doctors, which led to unfilled items or missing values in the dataset. The higher the proportion of missing data is, the more difficult it is to restore the information of the original data, and the more likely it is to affect the performance of the model. Generally, discarding a variable is considered when the missing rate exceeds 50% because it may not have significant value for analysis; if the missing rate of a certain variable is below 50%, then data filling is considered.

In this study, group mean value was adopted as the filling method. Data were grouped according to the label category, and missing values within each group were then filled. The advantage of group filling is that it can maintain data consistency and integrity, avoiding additional noise or bias.

Model setting

The purpose of this study was to conduct multicenter study and realize the collaborative training of a diagnostic model for severe pediatric pneumonia based on the Arya privacy computing platform while protecting the data privacy of each hospital. The system consisted of a central server and four hospital clients. During each iteration of federated learning, the central server collected all local models uploaded by the hospital, calculated the global model of the current iteration through the federated average algorithm, and sent the global model to each hospital. Each hospital received the current global model from the server, used local data to continue training on the basis of the global model, and obtained a new local model.

For the machine learning algorithm, model training was performed via deep neural network with the following structure (Figure 3): 16 → 32 → 32 → 16 → 1. The activation functions of each layer were rectified linear unit (ReLU)→ ReLU → ReLU → ReLU → Sigmoid. The maximum number of iterations for federated learning was200, and the model was optimized with Adam algorithm with a learning rate of 0.05.

Figure 3 The structure of the developed model. ReLU, rectified linear unit.

Model training and evaluation

In order to verify the performance of federated learning technology on pneumonia detection, three groups were delineated.

Group 1 involved single-center modeling, with data from the Children’s Hospital, Zhejiang University School of Medicine used for single-center model training and evaluation. Group 2 involved two-center modeling, with data from the Children’s Hospital, Zhejiang University School of Medicine and one of the other three hospitals being selected for two-center model training via federated learning technology. The performance was evaluated with data from Children’s Hospital, Zhejiang University School of Medicine, and finally, the two-center model training was conducted three times. Group 3 involved four-center modeling, with all the data being selected for model training by federated learning technology to develop a global model for severe pneumonia detection. Its performance was evaluated with data from the Children’s Hospital, Zhejiang University School of Medicine.

In this study, the model efficacy was evaluated according to sensitivity, specificity, and accuracy. Sensitivity (or the true-positive rate) is the proportion of positive cases that are correctly predicted by a model. The higher the sensitivity is, the stronger the model’s ability to recognize positive cases. The formula for calculating sensitivity is as follows:

Sensitivity=TPTP+FN[1]

where true positive (TP) is the number of samples that are actually positive and predicted as positive and false negative (FN) is the number of samples that are actually positive but predicted as negative.

Specificity (or the true-negative rate) is the proportion of negative cases that are correctly predicted by the model. The higher the specificity is, the stronger the model’s ability to exclude negative cases. The formula for calculating specificity is as follows:

Specificity=TNTN+FP[2]

where true negative (TN) is the number of samples that are actually negative and predicted as negative, and false positive (FP) is the number of samples that are actually negative but predicted as positive.

Accuracy is the proportion of all cases that are correctly predicted by the model. The higher the accuracy is, the better the model’s overall performance. The formula for calculating accuracy is as follows:

Accuracy=TP+TNTP+TN+FP+FN[3]

The workflow of data preprocessing and model training

The overview of data preprocessing and model training is provided in Figure 4.

Figure 4 Workflow of data preprocessing and model training.

Results

Performance of single-center modeling

As shown in Table 2, on the training set, the single-center model achieved a sensitivity of 99.20%, a specificity of 83.00%, and an accuracy of 86.20 % in detecting severe pneumonia. On the test set, the single-center model achieved a sensitivity of 88.10%, a specificity of 74.60%, and an accuracy of 81.00% in detecting severe pneumonia.

Performance of two-center modeling

As shown in Table 2, on the training set, the two-center model based on the Children’s Hospital, Zhejiang University School of Medicine and the Children’s Hospital of Soochow University achieved a sensitivity of 99.00%, a specificity of 79.50%, and an accuracy of 86.40% in detecting severe pneumonia. On the test set, the model achieved a sensitivity of 90.80%, a specificity of 73.20%, and an accuracy of 82.70% in detecting severe pneumonia.

As shown in Table 2, on the training set, the two-center model based on the Children’s Hospital, Zhejiang University School of Medicine and the Changchun Children’s Hospital achieved a sensitivity of 100.00%, a specificity of 85.00%, and an accuracy of 87.20% in detecting severe pneumonia. On the test set, the model achieved a sensitivity of 94.60%, a specificity of 75.50%, and an accuracy of 82.00% in detecting severe pneumonia.

As shown in Table 2, on the training set, the two-center model based on the Children’s Hospital, Zhejiang University School of Medicine and the Kunming Children’s Hospital achieved a sensitivity of 98.80%, a specificity of 80.80%, and an accuracy of 85.50% in detecting severe pneumonia. On the test set, the model achieved a sensitivity of 94.10%, a specificity of 79.20%, and an accuracy of 83.20% in detecting severe pneumonia.

Performance of four-center modeling

As shown in Table 2, on the training set, the four-center model achieved a sensitivity of 100.00%, a specificity of 85.30%, and an accuracy of 85.00% in detecting severe pneumonia. On the test set, the model achieved a sensitivity of 95.10%, a specificity of 82.70%, and an accuracy of 85.80% in detecting severe pneumonia.

Discussion

In this study, we employed the Arya privacy computing platform to conduct a multicenter retrospective study and develop several diagnostic models for severe pneumonia in pediatric patients. To our knowledge, this is the first work of its kind to apply a privacy computing plat form to a multicenter study with a large sample size to construct diagnostic models for severe pediatric pneumonia (a total of 5,091 patients with pediatric pneumonia). The privacy computing platform facilitated the performance of a multicenter study and aided in developing a reliable diagnostic model for severe pneumonia, with no transfer of data being required.

Although several diagnostic models on severe pediatric pneumonia have been developed with good discrimination and calibration (7,8), their efficacy has been not been validated in multicenter studies with a large sample. We conducted a multicenter retrospective study with a sample of 5,091 patients with pediatric pneumonia. Moreover, to avoid the disclosure of sensitive medical information to researchers, other institutions, and unauthorized users, we applied the Arya privacy computing platform. This platform (14) was developed by Hangzhou Healink Technology Corporation, who integrated the latest technologies, including federated learning, secure multiparty computation, distributed machine learning, and block chain to balance data value and privacy protection. This innovation can help medical centers better apply data to conduct multicenter studies while ensuring the security of the original data.

Four hospitals participated into the study. During model training, Arya helped to build a single data platform with each hospital as the core. A data collaboration alliance network was then built, which provided two-way promotion. In addition, in federated learning technology, each participant can only exchange encrypted model parameters and complete the establishment of the model without exchanging or transferring original data (15-17). To verify the efficacy of a multicenter study in training a diagnostic model for severe pneumonia, we validated and compared the diagnostic efficacy of models developed with datasets from one center, two centers, and four centers. We found that the incorporation of privacy-preserving federated learning technology into a multicenter study enabled the development of models with reliable performance in diagnosing severe pneumonia. Among these models, the four-center model achieved the highest diagnostic efficacy (95.10% sensitivity, 82.70% specificity, and 85.80% accuracy). Although the two-center models yielded a relatively low diagnostic efficacy, they still surpassed the diagnostic efficacy of the single-center model (88.10% sensitivity, 74.60% specificity, and 81.00% accuracy). These findings indicate that a multicenter model provides superior diagnostic efficacy.

Certain limitations to this study should be addressed. First, the sample size varied significantly across centers, potentially introducing statistical bias. Second, only 11 of 33 collected indicators were used due to inconsistent data quality across centers, which might have limited the model’s comprehensiveness. Third, the retrospective design relies on historical data, which may not fully reflect current diagnostic practices. Concerns regarding data security have long been an obstacle to the development of multicenter research (18,19). Our study serves as an example of the successful application of a privacy computing platform to multicenter research. In the future, data security technology will allow for the broader application of data-driven studies in clinical research—especially those involving multiple centers—strengthen the collaboration between institutions, promote new discoveries, and accelerate the translation of laboratory findings into clinical practice.

Conclusions

Privacy-preserving federated learning technology can facilitate the performance of multicenter studies and was used to develop a high-performance diagnostic model for severe pneumonia in pediatric patients, which can benefit doctors and patients as an auxiliary diagnostic tool.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-349/rc

Data Sharing Statement: Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-349/dss

Peer Review File: Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-349/prf

Funding: This work was supported by the Key Research and Development Program of Zhejiang (grant No. 2023C03101).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-349/coif). G.Y. serves as an unpaid editorial board member of Translational Pediatrics. The other 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. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This multicenter retrospective study was approved by the Medical Ethics Committee of Children’s Hospital, Zhejiang University School of Medicine (No. 2023-IRB-0181-P-01) and the requirement for written informed consent from patients was waived as long as the data of the patients’ remained anonymous. All participating hospitals/institutions were informed and agreed with the 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. Williams DJ, Zhu Y, Grijalva CG, et al. Predicting Severe Pneumonia Outcomes in Children. Pediatrics 2016;138:e20161019. [Crossref] [PubMed]
2. Liu F, Chen L, Wang MY, et al. Exploring high-risk factors for the prediction of severe mycoplasma pneumonia in children. Transl Pediatr 2024;13:2003-11. [Crossref] [PubMed]
3. Wang N, Fang Y, Dong H, et al. Clinical features and prediction of risk factors for severe adenovirus pneumonia in children. Transl Pediatr 2024;13:63-71. [Crossref] [PubMed]
4. Ding G, Zhang X, Vinturache A, et al. Challenges in the treatment of pediatric Mycoplasma pneumoniae pneumonia. Eur J Pediatr 2024;183:3001-11. [Crossref] [PubMed]
5. Renaud B, Coma E, Labarere J, et al. Routine use of the Pneumonia Severity Index for guiding the site-of-treatment decision of patients with pneumonia in the emergency department: a multicenter, prospective, observational, controlled cohort study. Clin Infect Dis 2007;44:41-9. [Crossref] [PubMed]
6. Rueda ZV, Aguilar Y, Maya MA, et al. Etiology and the challenge of diagnostic testing of community-acquired pneumonia in children and adolescents. BMC Pediatr 2022;22:169. [Crossref] [PubMed]
7. Chang Q, Chen HL, Wu NS, et al. Prediction Model for Severe Mycoplasma pneumoniae Pneumonia in Pediatric Patients by Admission Laboratory Indicators. J Trop Pediatr 2022;68:fmac059. [Crossref] [PubMed]
8. Haggie S, Barnes EH, Selvadurai H, et al. Paediatric pneumonia: deriving a model to identify severe disease. Arch Dis Child 2022;107:491-6. [Crossref] [PubMed]
9. Li M, Wei X, Zhang SS, et al. Recognition of refractory Mycoplasma pneumoniae pneumonia among Myocoplasma pneumoniae pneumonia in hospitalized children: development and validation of a predictive nomogram model. BMC Pulm Med 2023;23:383. [Crossref] [PubMed]
10. Katreddi S, Midatani A, Roy AP, et al. Pediatric pneumonia X-ray image classification: predictive model development with DenseNet-169 transfer learning. J Med Artif Intell 2025;8:37.
11. Son Y, Han K, Lee YS, et al. Privacy-preserving breast cancer recurrence prediction based on homomorphic encryption and secure two party computation. PLoS One 2021;16:e0260681. [Crossref] [PubMed]
12. Samaga D, Hornung R, Braselmann H, et al. Single-center versus multi-center data sets for molecular prognostic modeling: a simulation study. Radiat Oncol 2020;15:109. [Crossref] [PubMed]
13. Bleker J, Yakar D, van Noort B, et al. Single-center versus multi-center biparametric MRI radiomics approach for clinically significant peripheral zone prostate cancer. Insights Imaging 2021;12:150. [Crossref] [PubMed]
14. Xu J, Zhang Y, Yu H, et al. High performance of privacy-preserving acute myocardial infarction auxiliary diagnosis based on federated learning: a multicenter retrospective study. Ann Transl Med 2022;10:1006. [Crossref] [PubMed]
15. Simm J, Humbeck L, Zalewski A, et al. Splitting chemical structure data sets for federated privacy-preserving machine learning. J Cheminform 2021;13:96. [Crossref] [PubMed]
16. Wang R, Lai J, Zhang Z, et al. Privacy-Preserving Federated Learning for Internet of Medical Things Under Edge Computing. IEEE J Biomed Health Inform 2023;27:854-65. [Crossref] [PubMed]
17. Peyvandi A, Majidi B, Peyvandi S, et al. Privacy-preserving federated learning for scalable and high data quality computational-intelligence-as-a-service in Society 5.0. Multimed Tools Appl 2022;81:25029-50. [Crossref] [PubMed]
18. Dalrymple HW. The general data protection regulation, the clinical trial regulation and some complex interplay in paediatric clinical trials. Eur J Pediatr 2021;180:1371-9. [Crossref] [PubMed]
19. Mello MM, Lieou V, Goodman SN. Clinical Trial Participants' Views of the Risks and Benefits of Data Sharing. N Engl J Med 2018;378:2202-11. [Crossref] [PubMed]

(English Language Editor: J. Gray)