A Cohort Study of Pediatric Severe Community-Acquired Pneumonia Involving AI-Based CT Image Parameters and Electronic Health Record Data

Abstract

Introduction

Community-acquired pneumonia (CAP) is a significant concern for children worldwide and is associated with a high morbidity and mortality. To improve patient outcomes, early intervention and accurate diagnosis are essential. Artificial intelligence (AI) can mine and label imaging data and thus may contribute to precision research and personalized clinical management.

Methods

The baseline characteristics of 230 children with severe CAP hospitalized from January 2023 to October 2024 were retrospectively analyzed. The patients were divided into two groups according to the presence of respiratory failure. The predictive ability of AI-derived chest CT (computed tomography) indices alone for respiratory failure was assessed via logistic regression analysis. ROC (receiver operating characteristic) curves were plotted for these regression models.

Results

After adjusting for age, white blood cell count, neutrophils, lymphocytes, creatinine, wheezing, and fever > 5 days, a greater number of involved lung lobes [odds ratio 1.347, 95% confidence interval (95% CI) 1.036–1.750, P = 0.026] and bilateral lung involvement (odds ratio 2.734, 95% CI 1.084–6.893, P = 0.033) were significantly associated with respiratory failure. The discriminatory power (as measured by the area under curve) of Model 2 and Model 3, which included electronic health record data and the accuracy of CT imaging features, was better than that of Model 0 and Model 1, which contained only the chest CT parameters. The sensitivity and specificity of Model 2 at the optimal critical value (0.441) were 84.3% and 59.8%, respectively. The sensitivity and specificity of Model 3 at the optimal critical value (0.446) were 68.6% and 76.0%, respectively.

Conclusion

The use of AI-derived chest CT indices may achieve high diagnostic accuracy and guide precise interventions for patients with severe CAP. However, clinical, laboratory, and AI-derived chest CT indices should be included to accurately predict and treat severe CAP.

Supplementary Information

The online version contains supplementary material available at 10.1007/s40121-025-01197-0.

Key Summary Points

Why carry out this study?

Currently artificial intelligence (AI) and radiomics are feasible and well-performing tools to assist clinical decisions in the diagnosis, prognosis, and assessment of lung disease.

Community-acquired pneumonia (CAP) in children is a significant concern associated with high morbidity and mortality worldwide.

However, studies of severe CAP prognosis using AI-based computed tomography image analysis on pediatric patients has been scarce thus far.

What was learned from the study?

The combination of AI and chest CT images may enhance the accuracy of diagnostics and helpfully offer precise interventions for severe CAP in pediatric patients.

It is necessary to combine CT imaging features, clinical features, and demographic information to predict respiratory failure in pediatric patients with CAP.

This would allow clinicians to precisely identify children who are at risk of severe CAP and begin early therapies or preventative steps to enhance outcomes.

Introduction

Community-acquired pneumonia (CAP) is an acute lung infection that occurs in the extra-hospital setting with high healthcare costs and is a major cause of childhood morbidity and mortality worldwide, particularly among children younger than 5 years [1, 2]. A global, regional, and national incidence and mortality burden study revealed that there were approximately 344 million new episodes of CAP, or 4350 episodes per 100,000 individuals, and 2.18 million deaths, or 27.7 deaths per 100,000 individuals, in 2021 [3]. Severe CAP is the most life-threatening form of CAP [4] and is associated with a large number of complications, among which respiratory failure is one of the most common and severe [5–7].

Currently, exciting advancements in artificial intelligence (AI) are leading new developments in medical imaging analysis. AI, especially machine learning, can also improve the efficiency of CAP prediction, diagnosis, and management by analyzing large-scale real-world data [8]. An international multicenter, multivendor computed tomography(CT) study found that the diagnostic accuracy of eight radiologists was 79.1% without AI assistance and increased to 81.5% with AI assistance, accompanied by significantly shorter decision times and higher confidence scores in CAP [9]. Shakibfar et al. also developed a disease risk score based on population-wide real-world data and showed that it had an average accuracy of 0.79 based on fivefold cross-validation in predicting CAP hospitalization [10]. Moreover, AI algorithms reduce variability associated with subjective clinical judgment, promoting consistency in decision-making, and provide real-time risk assessments, aiding in the dynamic management of patients with CAP [11]. Furthermore, AI and radiomics have been shown to be feasible and well-performing tools for assisting in clinical decision-making in the diagnosis, prognosis, and assessment of lung disease [12, 13]. Shao et al. reported that the resolution index of an AI-based multimodal integration system comprising clinical text and CT images in detecting pulmonary infections was 91.0%, suggesting that the system can provide advantages in predicting the risk of developing critical illness and contribute to more informed clinical decision-making [12]. Additionally, a deep learning-based medical image interpretation system, which could provide specific parameters of CT, achieved an average area under the receiver operating characteristic curve (AUG) of 0.90 in diagnosing major respiratory diseases on the basis of CT images [14]. Nevertheless, these studies on the outcomes of pediatric patients with severe CAP have been scarce thus far.

This study sought to evaluate the predictive value of AI-based CT image parameters and electronic health record data for pediatric patients with CAP. This would allow clinicians to precisely identify children who are at risk of severe CAP and begin early therapies or preventative steps to enhance outcomes. We hypothesized that chest CT parameters had a statistically significant correlation with acute respiratory failure in children with severe CAP.

Methods

Study Population

Pediatric patients who were hospitalized for severe CAP at the Seventh Affiliated Hospital of Sun Yat-Sen University in Shenzhen from January 2023 to October 2024 were enrolled in this retrospective cohort study. This study was approved by the Medical Ethics Committee of the Seventh Affiliated Hospital of Sun Yat-Sen University (approval number KY-2024-312-01). Given the noninterventional and retrospective nature of the study, the ethics committee waived the need for written informed consent from either the patients or their legal guardians. This study was performed in accordance with the Helsinki Declaration of 1964 and its later amendments to the ethics statement.

The inclusion criteria for the population were as follows: (1) diagnosis of severe CAP according to the Guidelines for the Management of Community-Acquired Pneumonia in Children (2024 revision) from the National Health Commission of the People’s Republic of China [15]; (2) evidence consistent with acute respiratory failure [16, 17]: hypoxemic without hypercapnia and hypercapnic acidosis (hypoxemia is characterized by PaO₂ < 60 mmHg on ambient air, SpO₂ < 90% on ambient air or the need to administer oxygen to achieve PaO₂ ≥ 60 mmHg or SpO₂ ≥ 90%. Hypercapnic acidosis is characterized by pH ≤ 7.35, with PaCO₂ > 45 mmHg), or PaO₂/FiO₂ ratio of ≤ 300 mmHg; (3) postadmission chest CT data available; and (4) age under 18 years.

The exclusion criteria were as follows: (1) diseases such as lung tumors, severe malnutrition, bronchiectasis, chronic cardiac, congenital disease, or tuberculosis, or hematopoietic stem cell transplantation within 90 days; (2) no chest CT data; (3) an average Likert score < 3 as evaluated by two radiologists or CT scans that did not meet the analytical criteria because of respiratory artifacts [18]; (4) discharge from the hospital without consulting the treating physician; and (5) death.

Chest CT Images

All patients underwent non-enhanced chest CT imaging with a Siemens SOMATOM Force CT scanner (Siemens Healthcare, Forchheim, Germany) at our hospital, with the scan covering from the apex to the base of the lungs. The image parameters were as follows [19]: field of view (FOV), 300 × 300 mm; detector collimation, 2 × 192 × 0.6; rotation time, 0.25 s; voltage, 100 kV; current, 60 mAs; and section thickness, 1 mm.

CT Images Analysis

AI software [Computer-Aided Analysis System for Pneumonia (uAI Discover-Pneumonia, Shanghai United Imaging Intelligence Co. Ltd.)] was used to obtain CT image parameters, including the number of involved lung lobes, bilateral lung involvement, proportion of lung infection, and volume of infected lung tissue from the patients’ CT images. This algorithm consists of three modules: (a) a pneumonia diagnostic module, (b) a pulmonary lobe infection module, and (c) a pulmonary segment infection module.

Two skilled radiologists (JP Yuan and JS Yi) independently and visually assessed all the segmentation results produced by the deep learning algorithm. A Likert scale ranging from 0 to 5 points was used to describe the degree of matching [18]. A score of zero was given if the radiologist identified either medium-sized regions of false-positive or false-negative contours on at least three slices or very large regions of false-positive or false-negative contours on at least one slice. If at least one slice had a medium-sized region of false-positive or false-negative contours, a score of 1 was given if the requirements for a score of 0 were not reached. A score of 3 indicated that every slice had no discernible false-positive or false-negative features. A score of 5 indicated that all slices were segmented perfectly along the actual lung opacification. Intermediate circumstances that did not fit the previously stated requirements were given scores of 2 and 4. The final score was the average of the two scores for each scan to lessen the influence of the subjectivity of the radiologist’s assessments on the score. A final score of ≥ 3 was deemed adequate to fulfill the criteria for quantitative analysis. Cohen’s kappa coefficient for agreement between the two radiologists in their judgments was 0.665 (P < 0.001).

Data Collection

For all enrolled patients, the following data were gathered either through use of the AI software or from their electronic medical records: (1) demographic factors such as body mass index, sex, and age; (2) chest CT indices, including the number of involved lung lobes, bilateral lung involvement, proportion of lung infection, and volume of lung infection; (3) laboratory parameters, including white blood cell count and the levels of C-reactive protein, lactate procalcitonin, dehydrogenase, alanine aminotransferase, and aspartate aminotransferase; and (4) clinical symptoms and physical signs, including fever, cough, and wheezing.

Statistical Analysis

Continuous variables are presented as the means ± standard deviations or medians (interquartile ranges) and were compared with paired the t test or Wilcoxon rank-sum test. Numbers and percentages are reported for categorical variables, which were compared with the chi-square test. Using two independent sample t test (α = 0.05, power = 90%; effect size: OR1 (odds ratio) = 2.35 [20], P2 = 0.59 [21]) in PASS, version 15.0, each group required 130 participants. Multiple imputation by chained equations, which adjusted for all measured variables potentially associated with missing data, was used to impute missing clinical data [22]. Logistic regression analysis was used to determine whether the chest CT indices could be used alone to predict respiratory failure. A multifactorial regression model was constructed on the basis of baseline variables that were clinically significant or yielded a P value < 0.2 in the univariable analyses. To ensure that the final model was easy to understand, the included variables were carefully selected. Receiver operating characteristic (ROC) curves of the indices in predicting respiratory failure were constructed. Statistical analysis was performed with IBM SPSS Statistics (version 25.0; IBM Corporation, Armonk, NY, USA). A statistically significant result was indicated for P < 0.05.

Results

On the basis of the inclusion and exclusion criteria, 230 individuals were ultimately recruited into this study, including 51 (22.2%) with respiratory failure (RF) and 179 (77.8%) without respiratory failure (NRF), as shown in Fig. 1.

Fig. 1.

Flowchart of the study population. CT computed tomography

Baseline Characteristics of the Study Population

The patients with RF had a median age of 3.6 (1.9, 5.9) years, 70.6% of them were ≤ 5 years of age, and 47.1% were male. The patients with NRF had a median age of 4.7 (2.3, 6.6) years, 57.0% of them were ≤ 5 years of age, and 50.8% were male. The two groups differed considerably in the proportion with symptoms of wheezing, lymphocyte percentage, neutrophils percentage, creatinine level, volume and proportion of lung infection, number of implicated lung lobes, and bilateral lung involvement (P < 0.05) (Table 1). However, there was no difference in other laboratory characteristics, radiological features, or clinical symptoms between the two groups (P > 0.05).

Table 1.

Clinical characteristics

	RF (n = 51)	NRF (n = 179)	P
Age, years	3.6 (1.9, 5.9)	4.7 (2.3, 6.6)	0.094
≤ 5 years	36 (70.6%)	102 (57.0%)	0.080
Male	24 (47.1%)	91 (50.8%)	0.634
Female	27 (52.9%)	88 (49.2)
BMI, kg/m2	15.5 (14.8, 16.9)	15.2 (14.1, 16.3)	0.231
Fever	41 (80.4%)	150 (83.8%)	0.567
> 5 days	4 (7.8%)	34 (19.0%)	0.059
Cough	51 (100%)	179 (100.00%)	–
Wheezing	24 (47.1%)	47 (26.3%)	0.005*
White blood cell, × 109/L	9.2 (6.2, 13.4)	7.5 (6.0, 10.5)	0.067
Lymphocytes, %	20.7 (13.2, 33.0)	28.4 (22.5, 41.6)	0.001*
Neutrophils, %	70.2 (55.3, 79.6)	60.9 (49.2, 69.5)	0.001*
CRP ≥ 10 mg/L	31 (60.8%)	97 (54.2%)	0.403
PCT, ng/mL
< 0.05	15 (29.4%)	55 (30.7%)	1.000
[0.05, 2)	34 (66.7%)	118 (65.9%)
≥ 2	2 (3.9%)	6 (3.4%)
ALT, U/L	18.1 (14.6, 24.9)	18.1 (14.9, 23.1)	0.999
AST, U/L	36.4 (31.8, 41.1)	37.3 (30.6, 45.6)	0.478
LDH, U/L	289.8 (270.9, 353.7)	306.1 (269.8, 362.6)	0.459
d-D > 5 times the upper limit of normal	3 (5.9%)	7 (3.9%)	0.465
Urea, mmol/L	3.4 (2.5, 3.9)	3.3 (2.5, 4.0)	0.949
Creatinine, μmol/L	27.3 (23.0, 32.9)	30.6 (24.2, 35.4)	0.038*
Proportion of lung infection, %	3.2 (1.3, 10.6)	6.6 (1.9, 13.1)	0.047*
Volume of lung infection, cm3	25.3 (7.5, 74.1)	49.8 (18.0, 107.6)	0.026*
Number of involved lung lobes	5 (3, 5)	4 (2, 5)	0.035*
Pulmonary consolidation	42 (82.4%)	142 (79.3%)	0.634
Pleural effusion	4 (7.8%)	16 (8.9%)	1.000
Bilateral lung involvement	44 (86.3%)	128 (71.5%)	0.032*

RF respiratory failure, NRF no respiratory failure, BMI body mass index, CRP C-reactive protein, PCT procalcitonin, AST alanine aminotransferase, ALT aspartate aminotransferase, LDH lactic dehydrogenase, d-D d-dimer

*Statistically significant (P < 0.05)

Chest CT Indices and Respiratory Failure

Binary regression analysis was performed to assess whether the chest CT indices were associated with respiratory failure and constructed on the basis of baseline variables that were clinically significant or yielded a P value < 0.2 in the univariable analyses. To ensure that the final model was easy to understand, the variables finally included age, white blood cell count, lymphocytes, creatinine, wheezing, and fever > 5 days. We added potential multicollinearity analysis among these variables and found variance inflation factors all were < 5 (Table S1), which indicated that there was no collinearity among these variables.

After adjusting for age, white blood cell count, lymphocytes, creatinine, wheezing, and fever > 5 days, a greater number of involved lung lobes (odds ratio 1.347, 95% CI 1.036–1.750, P = 0.026) and bilateral lung involvement (odds ratio 2.734, 95% CI 1.084–6.893, P = 0.033) were significantly associated with respiratory failure, as shown in Table 2.

Table 2.

Chest computed tomography indices and respiratory failure (see Table S2 for further information)

	OR	P	95% CI
			Lower	Upper
Number of involved lung lobes	1.347	0.026	1.036	1.750
Bilateral lung involvement	2.734	0.033	1.084	6.893
Proportion of lung infection	0.991	0.639	0.955	1.028
Volume of lung infection	0.997	0.388	0.991	1.003

OR odds ratio, CI confidence interval

Figure 2 shows the ROC curves of these variables in predicting respiratory failure. Model 2: logit(P) = − 0.180 + 0.298 × number of involved lung lobes − 0.020 × creatinine level − 0.052 × lymphocyte percentage + 0.032 × white blood cell count + 0.515 × wheezing − 0.481 × fever > 5 days − 0.132 × age. The ROC curve shows that the AUG for predicting respiratory failure in 230 cases of severe CAP based on Model 2 is 0.763 (95% CI 0.691–0.835), which was better than that of Model 0 (AUG = 0.593, 95% CI 0.506–0.681), which contained only the number of lung lobe infections. The sensitivity and specificity of Model 2 at the optimal critical value (0.441) were 84.3% and 59.8%, respectively. Model 3: logit(P) = 0.160 + 1.006 × bilateral lung involvement − 0.021 × creatinine level − 0.050 × lymphocyte percentage + 0.035 × white blood cell count + 0.439 × wheezing − 0.506 × fever > 5 days − 0.139 × age. The ROC curve shows that the AUG for predicting respiratory failure based on Model 3 is 0.763 (95% CI 95% CI 0.690–0.835), which is better than the single indicator of bilateral lung involvement (Model 1: AUG = 0.574, 95% CI 0.489–0.658). The sensitivity and specificity of Model 3 at the optimal critical value (0.446) were 68.6% and 76.0%, respectively.

Fig. 2.

Areas under the receiver operating characteristic (ROC) curve for respiratory failure. Model 0: Number of involved lung lobes and respiratory failure. Model 1: bilateral lung involvement and respiratory failure. Model 2: number of involved lung lobes, age, white blood cell count, lymphocyte percentage, creatinine level, wheezing, and fever > 5 days; Model 3: bilateral lung involvement, age, white blood cell count, lymphocyte percentage, creatinine level, wheezing, and fever > 5 days. AUG area under curve

Discussion

In this cohort study, we examined the AI-based CT image parameters, clinical features, and demographic information as combined in four different models to predict respiratory failure in pediatric patients with severe CAP. The discriminatory power (as measured by the AUG) of Model 2 and Model 3, which included electronic health record data and the accuracy of CT imaging features, was better than that of Model 0 and Model 1, which contained only the chest CT parameters. This may help risk stratify for severe CAP in hospitalized children and determine which patients require inpatient or transfer to the pediatric intensive care unit.

CAP is a leading cause of hospitalization and death in children, and one of its most severe complications is respiratory failure. Radiologic imaging plays an essential role in the screening, diagnosis, and outcome prediction of patients with a variety of respiratory diseases [14, 23, 24]. Although chest X-ray is the most commonly used first-line radiological method for assessing respiratory diseases, chest CT can provide more accurate information for clinicians and is a currently key component of medical imaging techniques for the identification of CAP [25]. Therefore, we analyzed the CT images of patients with CAP in our study.

Draelos et al. [23] analyzed a CT data set of 36,316 volumes from 19,993 unique patients and developed a model for multiorgan, multi-disease classification based on chest CT volumes, which reached an AUG > 0.90 in classifying 18 abnormalities and an average AUG in identifying a total of 83 abnormalities, demonstrating the feasibility of machine learning from unfiltered, whole-volume CT data. Similarly, Hwang et al. [24] developed a deep learning-based automated detection algorithm for major thoracic diseases on the basis of chest radiographs and reported significant improvements in both image-wise classification and lesion-wise localization for all three physician groups with the assistance of the algorithm. Furthermore, a cluster analysis of thoracic muscle mass from chest CT scans using AI in severe pneumonia showed that higher muscle mass correlated with lower in-hospital mortality and improved clinical outcomes like extubation success and the model integrating muscle mass metrics outperformed conventional scores, with an AUG of 0.844 for predicting extubation success and 0.696 for predicting mortality [26]. In our study, we used a deep chest CT learning algorithm and found that two chest CT indices (number of involved lung lobes: odds ratio 1.347, 95% CI 1.036–1.750, P = 0.026; bilateral lung involvement: odds ratio 2.734, 95% CI 1.084–6.893, P = 0.033) were significantly associated with respiratory failure due to severe CAP. We thus surmised that AI-based CT imaging parameters have complementary value for prognostic prediction of severe CAP in children.

Chest radiograph interpretation is a difficult and error-prone procedure that requires skilled readers. Currently, these issues might be resolved with the use of automated systems that can correctly categorize images [27–29]. A study assessing the applicability and acceptance of a deep convolutional neural network (CNN)-based model by physicians reported that senior physicians had a high agreement rate (70%) with this system in identifying logical consolidation patterns and that three medical residents reached a higher agreement using support decision tool than alone with experts (0.66 ± 0.1 vs. 0.75 ± 0.2) [30]. Therefore, AI-based read assistants may hopefully aid clinicians in improving their efficiency and enhancing accuracy in identifying chest images.

However, precise disease diagnosis and treatment require integrated multimodal data analysis that does not rely solely on images but also includes clinical features, laboratory tests, and demographic information [31–34]. Jullien et al. confirmed that inflammatory and endothelial activation markers could be proactively used at first encounter as risk stratification and clinical decision-making tools in pediatric pneumonia [35]. A systematic review and meta-analysis also showed that serum creatinine, blood urea nitrogen, and C-reactive protein were significantly correlated with the risk of death from severe pneumonia [36]. A national, regional, and state-level study in India reported that age (such as < 5 years old) is a high-risk group for severe pneumonia, especially in low-income areas [37]. In our study, Model 2 and Model 3, which included demographic information (age), clinical complaints (fever and wheezing), laboratory test indicators (white blood cell count, lymphocyte percentage, and creatinine level), and chest imaging indices (number of involved lung lobes and bilateral lung involvement), demonstrated superior diagnostic accuracy for respiratory failure due to severe CAP over the CT image index-only model (AUGs 0.757 and 0.758 vs. 0.591, respectively). Similarly, Shao et al. [12] reported that a model consisting of unprocessed laboratory test indicators, clinical complaints, demographic information, and automatically coded and embedded chest images demonstrated greater diagnostic accuracy for pulmonary diseases than a single-dimensional model (AUG 0.887 vs. 0.840). In conclusion, despite the notable efficacy of the combination of AI and CT imaging, demographic characteristics, clinical features, laboratory examinations, and radiologic findings should all be considered in providing pediatric patients with accurate treatment. However, this paper did not test the acceptability and applicability of our model to families and clinicians. Further external validation is necessary.

There are certain limitations in our study that should be recognized. First, this was a single-center study, and the results may not have sufficient generalizability for all patients who experience respiratory failure caused by severe CAP. Multicenter studies with larger samples are needed. Additionally, we did not carry out prospective, external, or internal validation of our models. In future research, we will focus on validating these results with more diverse datasets to improve the generalizability of the AI system.

Conclusion

Chest CT indices (number of involved lung lobes and bilateral lung involvement) were significantly associated with respiratory failure due to severe CAP in pediatric patients. The combination of AI and chest CT images may achieve high diagnostic accuracy and help suggest precise interventions for patients with respiratory failure caused by severe CAP. Clinical features, laboratory tests, demographic information, and AI-derived chest CT indices should all be considered in the outcome prediction and treatment of patients with severe CAP.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contributions

Mengyuan He, Jianpeng Yuan, Aijiao Liu, Rui Pu, Wenqi Yu, Yinzhu Wang, Li Wang, Xing Nie, Jinsheng Yi, Hongman Xue and Junfeng Xie contributed to the study conception and design. Material preparation, data collection and analysis were performed by Mengyuan He, Junfeng Xie, Jianpeng Yuan and Hongman Xue. The first draft of the manuscript was written by Mengyuan He, Aijiao Liu and Jinsheng Yi. Mengyuan He, Jianpeng Yuan, Aijiao Liu, Rui Pu, Wenqi Yu, Yinzhu Wang, Li Wang, Xing Nie, Jinsheng Yi, Hongman Xue and Junfeng Xie commented on previous versions of the manuscript. Mengyuan He, Jianpeng Yuan, Aijiao Liu, Rui Pu, Wenqi Yu, Yinzhu Wang, Li Wang, Xing Nie, Jinsheng Yi, Hongman Xue and Junfeng Xie read and approved the final manuscript.

Funding

No funding or sponsorship was received for this study or publication of this article. The journal’s Rapid Service Fee was funded by Junfeng Xie.

Data Availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Conflict of Interest

Mengyuan He, Jianpeng Yuan, Aijiao Liu, Rui Pu, Wenqi Yu, Yinzhu Wang, Li Wang, Xing Nie, Jinsheng Yi, Hongman Xue and Junfeng Xie had nothing to disclose.

Ethical Approval

The Ethical Committee of the Seventh Affiliated Hospital of Sun Yat-Sen University approved the study (approval number KY-2024-312-01). Since this study was non-interventional and retrospective in nature, the ethical committee waived the need for written informed consent provided by patients or their legal guardians. This study was performed following the Helsinki Declaration of 1964 and its later amendments.