Introduction

Inflammatory bowel disease (IBD) is a complex and common chronic digestive tract disease, with approximately 20% of the cases presenting during childhood or adolescence. The incidence of pediatric IBD has been steadily increasing worldwide over the past decades, highlighting the urgent need for effective early diagnostic strategies.^1,2 Early diagnosis of pediatric IBD enables timely intervention, which can reduce disease complications, minimize growth and developmental delays, and improve long-term health outcomes.^3–5 However, in a large cohort recruited in a previous study, more than one-tenth of the patients exhibited symptoms at least 5 years before receiving a definitive diagnosis.⁶

The high heterogeneity of symptoms, inaccessibility of diagnostic instruments, and lack of specialized gastroenterologists and pathologists pose substantial barriers to the early identification of pediatric IBD.^7,8 Therefore, developing a reliable, time- and cost-effective diagnostic method to facilitate the early identification of children with IBD is urgently needed.

Although the Red Flag Index⁹ and IBD-REFER,¹⁰ developed based on clinical symptoms and routine laboratory test results, are currently available tools to help primary care physicians with early IBD detection, the former is not suitable for children and UC, and the latter has been widely validated internationally.

Intestinal ultrasonography is a noninvasive, radiation-free, and immediately available method for precise localization. Compared with the use of clinical symptoms and systemic inflammatory markers, the use of ultrasonography is more specific to the intestines and can localize lesions. Moreover, its value in the diagnosis and management of adults and children with IBD has been proven in many reports.^11–13 The accuracy of intestinal ultrasonography is inferior to that of small-bowel computed tomography enterography or small-bowel magnetic resonance imaging enterography.^14–16

Currently, few studies have assessed the combined use of ultrasonography with other noninvasive markers for the diagnosis of pediatric IBD. A systematic review published in 2019 reported that using endoscopy as the reference standard, the sensitivity of intestinal ultrasound, in whether detecting bowel wall thickening alone or in combination with other features (eg, vascularity, creeping fat, preserved wall stratification, and enlarged lymph nodes), in children ranged from 39% to 55%, with specificity ranging from 90% to 100%.¹⁷ In the study by DiLillo et al, the combination of bowel wall thickening and elevated fecal calprotectin (FC) improved specificity from 96.3% (88.5 to 99.1) to 100% (94.1 to 100), while maintaining a sensitivity of 69.6% (57.9 to 79.3), compared with bowel wall thickening alone.¹⁸

Following the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis¹⁹ we aimed to establish and validate the optimal model of noninvasive evaluation tests to help clinicians with the early identification of pediatric IBD.

Materials and Methods

Data Collection

Data for the training set were retrospectively collected from December 2021 to September 2023. This study targeted individuals who were newly diagnosed and referred, aged 2–18 years, and presenting with symptoms of suspected IBD, including recurrent abdominal pain, changes in bowel habits, hematochezia, and perianal abscess. The data was collected from electronic health records of the Children’s Hospital. The exclusion criteria were children diagnosed with known organic diseases (eg, previously identified IBD or celiac disease), those who had commenced IBD induction therapy, and those who had no available endoscopic or ultrasonographic data, or whose endoscopy or ultrasonography was performed more than one week apart (Figure 1). For the validation set, children were prospectively enrolled from January to June 2024, using the same inclusion criteria.

Figure 1 Flow diagram of the research participants. (a) Diagram of the flow of the training set. (b) Diagram of the flow of the validation set. (c) Diagnostic protocol for pediatric inflammatory bowel disease. Red-colored text indicates timing: (1) laboratory tests, endoscopy, and intestinal ultrasound were performed within one week; (2) the final reference standard diagnosis of IBD was made within one month following the completion of all assessments. Abbreviation: IBD, inflammatory bowel disease.

Outcome and Predictors

Our primary outcome of interest was a diagnosis of pediatric IBD. On the basis of previous diagnostic scales and literature,^9,10 as well as examinations conducted at our institution, we identified three sets of features. The first set included the following clinical symptoms: diarrhea lasting for at least 1 month, bloody stool lasting for at least 1 week, fever of an unknown cause, active perianal abscess, fistula or large inflamed skin tag, impaired growth, abdominal pain lasting for at least 1 month, weight loss (>10%), first-degree family history of IBD, rheumatological symptoms (ie, arthritis, uveitis, and erythema nodosum) not meeting a clear rheumatological diagnosis, and recurrent oral aphthous ulcerations. The second set included the following routine laboratory test results: elevated FC level, elevated C-reactive protein (CRP) level, elevated erythrocyte sedimentation rate (ESR), anemia, hypoalbuminemia, and positive antineutrophil cytoplasmic antibodies (ANCAs). The third set included the following intestinal ultrasonographic features: Limberg level >1 (bowel wall thickening with accompanying blood flow), mesenteric fat, disorganized or lost bowel wall layering, and enlarged lymph nodes.

Laboratory Evaluation

Laboratory tests included the complete blood count, albumin level, ESR, CRP level, and FC level. ANCA serotypes were assessed by indirect immunofluorescence with positive cytoplasmic ANCA or perinuclear ANCA and antigen-specific PR3 and MPO enzyme-linked immunosorbent assay. Blood count, ESR, and albumin values were determined using routine methods in the hospital laboratory. FC was measured using a quantitative immunochromatographic method, and the normal reference value is 0–50 μg/g. CRP was measured using the scattering turbidimetric method, and the normal reference value is less than 8 mg/L. Albumin was detected using the enzyme-linked immunosorbent assay, and the normal reference value was 39–54 g/L. ESR was detected using an ESR analyzer, and the normal reference value was 0–20 g/L. For ease of application, we converted the results for these markers into binary variables according to their normal cut-off values, similar to IBD-REFER.

Intestinal Ultrasound Examination

The same ultrasound expert performed all intestinal ultrasound examinations (>200 intestinal ultrasound examinations per year for 2 years). The ultrasound operator was aware of the patient’s symptoms and signs but was blinded to the endoscopic and laboratory findings. A Mindray R9 pro ultrasound system was used for the examination. A convex probe with a 1–7-MHz frequency was used for abdominal panoramic examination, followed by a high-frequency examination (3–12 MHz) using a linear probe for detailed intestinal evaluation. Color power Doppler imaging was applied, with the Doppler velocity range set to 3–5 cm/s. For each bowel segment, one image showing the maximal bowel wall thickness and one image capturing the area with the most prominent vascularity were saved as DICOM files. To assess inter-rater reliability, the primary operator and a second trained sonographer, who had completed institutional IUS training and had independently performed IUS in more than 50 pediatric IBD cases, independently scored the same set of saved images. Agreement was quantified using weighted kappa statistics.

To determine the maximum bowel wall thickening of each intestinal segment, we measured the longitudinal and transverse sections of the entire intestinal segment, excluding the rectum and duodenum. Mesenteric fat was characterized by the presence of hyperechoic areas extending from the serosa to the mesentery and effectively enveloping the outer bowel walls. Loss of echo-stratification was defined as when the multilayered bowel wall becomes prevalently hypoechogenic or hyperechogenic and is sometimes indiscernible.²⁰ Enlarged lymph nodes were defined as those with a short axis >5 mm or >5 planar lymph nodes (Supplementary Figure 1).

We employed the Limberg scoring system in conjunction with the assessments of intestinal wall thickness and blood flow signals. The Limberg scores^21,22 were categorized as follows: Limberg level 0 indicated no thickening of the intestinal wall; Limberg level 1 represented thickening of the intestinal wall without accompanying blood flow signals; Limberg level 2 denoted thickening of the intestinal wall (bowel wall thickening >3 mm) with punctate blood flow; Limberg level 3 indicated thickening of the intestinal wall with linear blood flow; and Limberg level 4 represented thickening of the intestinal wall with blood flow extending into the mesentery.

Definition of Inflammatory Bowel Disease

All enrolled participants had undergone endoscopy and had mucosal pathology examination data; some had capsule endoscopy data. A skilled pediatric gastroenterologist performed the endoscopy using Fuji devices. One month after follow-up, pediatric gastroenterologists determined whether the patients had IBD on the basis of the Porto criteria²³ while considering each patient’s endoscopy and pathology results and their follow-up conditions.

Assessment of Disease Activity

In patients with CD, clinical disease activity is assessed using the Pediatric Crohn’s Disease Activity Index (PCDAI). Endoscopic activity in CD is evaluated using the Simple Endoscopic Score for Crohn’s Disease (SESCD). For patients with UC, the Pediatric Ulcerative Colitis Activity Index (PUCAI) assesses clinical disease activity. Endoscopic activity in UC is measured using the Mayo Endoscopic Subscore (MES). Histological disease activity in all patients is evaluated using the Nancy histological index.²⁴

Development of the Model

Predictors were assessed by univariate logistic regression analysis, followed by an age-, sex- and disease duration-adjusted multivariate logistic regression analysis. To avoid false positives, a significance level of 1% was used as the threshold for confirming independent associated features. The effect sizes of significantly associated features were reported as odds ratios (OR) with their corresponding 95% confidence intervals (CI). Significant features were simultaneously subjected to LASSO regression and random forest analysis. LASSO regression, with an L1 penalty, was used to shrink the coefficients of less important variables to zero, thereby selecting key predictors while preventing overfitting. The optimal lambda value was determined through cross-validation. Meanwhile, random forest analysis was performed to rank features based on importance metrics for classification tasks, including accuracy decrease and Gini decrease. These importance metrics were calculated using the formulas provided in the “randomForest” and “randomForestExplainer” packages.^25,26 Finally, the intersection of significant features identified by both LASSO regression and random forest analysis was selected for logistic regression modeling.

Model Evaluation

To assess the model’s quality of fit, we employed the Hosmer–Lemeshow test and coefficient of determination (R2). To evaluate predictive accuracy and model conformity, we used the receiver operating characteristic curve, area under the receiver operating characteristic curve (AUC), concordance index, and calibration curve. Additionally, a decision curve analysis demonstrated the net benefit of the model for patients. Both discrimination and calibration were assessed through bootstrapping in the training dataset, involving 1000 resamples, and validation dataset. We also validated the IBD-REFER using the training dataset for direct comparison.

Spearman correlation coefficients were used to examine the relationships between predictive probabilities and disease severity (clinical, endoscopic, and histological). PCDAI, SESCD, PUCAI, MES, and Nancy scores were converted into ordinal variables to evaluate their correlation with the probability of IBD classification across IBD populations.

Statistical Analysis

Statistical analyses were performed using R (version 4.4.1; R Core Team 2024). Data are reported as frequencies or mean±standard deviation. The statistical significance of comparisons between groups was defined as <0.05. Continuous variables were tested for normality using the Kolmogorov–Smirnov test. Normally distributed variables were evaluated using t-tests. Dichotomous variables were evaluated against mortality using the chi-square or Fisher exact test. We calculated the sensitivity, specificity, positive predictive value, negative predictive value, and accuracy of the diagnostic methods using the formulas provided in the “caret” package.

Only the FC variable in the training dataset had 1% missing values, which were imputed using the mean of the FC values. The research location, inclusion and exclusion criteria, outcome measures, and predictive indicators of the validation and development datasets were similar.

Results

General Characteristics

In the final outcome assessment, 314 participants were included in the training dataset and 66 participants were included in the validation dataset. Table 1 shows the demographics of patients in the training dataset; the time from first symptoms to diagnosis in 74 (71.8%) children with IBD was >2 months. In the training dataset, 103 individuals were diagnosed with IBD, of which 89 had Crohn disease (CD), 13 had ulcerative colitis (UC), and one had IBD-Unclassified (Table 1). In the validation dataset, 19 individuals were diagnosed with IBD, of which 15 had CD and four had UC. The remaining 211 patients in the training dataset and 47 patients in the validation dataset were classified as non-IBD cases. The non-IBD group predominantly comprised individuals diagnosed with functional gastrointestinal disorders and allergic enteritis (Supplementary Table 1). For the same stored ultrasound images, the inter-rater agreement between two pediatric sonographers on Limberg grading was high, with a weighted kappa coefficient of 0.91, indicating excellent consistency.

Table 1 Clinical, Laboratory, and Ultrasonographic Characteristics of Training Dataset

Development of an Individualized Prediction Model Using the Training Dataset

Univariate analysis ruled out the following irrelevant factors: age, sex, abdominal pain, and first-degree family history (P>0.05). Fifteen factors remained significant after adjusting for age, sex and disease duration (P<0.01; Table 2). Figure 2a presents the results of the LASSO regression, identifying Limberg level >1, ESR, FC, CRP, hypoalbuminemia, and active perianal abscess or fistula as significant features. Similarly, Figure 2b shows the top features ranked by the random forest model based on Gini and accuracy decrease importance, which align with the LASSO results. Finally, these selected features were used to construct the logistic regression model. Diagnostic accuracy analysis was performed on these five features, with Limberg>1 demonstrating the highest accuracy (0.85). In the diagnostic interval subgroups, although the accuracy of ultrasound decreased slightly in the short diagnostic interval (<2 months) compared to the long diagnostic interval (≥2 months) (0.87 vs 0.83), it still maintained strong performance (Supplementary Figure 2). These features were then used to construct a logistic regression model. A personalized nomogram model was developed to categorize individuals according to their IBD risk scores (Figure 2b). By using the coefficients obtained from regression logistic analysis and applying the following formula, we calculated the estimated probability of children with IBD as 0.999:

Table 2 Univariate and Multivariate Logistic Regression Models for Diagnosing Pediatric IBD

Figure 2 Visualization of the diagnostic prediction models. (a) Evolutionary selection in LASSO regression. (b) Feature importance for the random forest model based on datasets with ultrasound characteristics. (c) Nomogram for diagnosing pediatric IBD. Abbreviations: IBD, inflammatory bowel disease; CRP, C-reactive protein; FC, fecal calprotectin; ESR, erythrocyte sedimentation rate; AUC, area under the curve; CI, confidence interval.

Logit (Probability)=−5.3092+2.7526×1 (Limberg level>1=Yes) +1.8198×1(elevated ESR = Yes) +2.7163×1 (elevated FC=Yes) +0.5974×1 (elevated CRP=Yes) +2.4585×1 (active perianal abscess or fistula=Yes) +2.0733×1 (hypoalbuminemia=Yes).

A free web-based calculator based on our model is available at: https://whlmt123456.github.io/PIBDcalculator.

Demonstrated Efficacy of the Nomogram in the Training and Validation Datasets

This nomogram of the training dataset exhibited a notable discriminative capability, achieving AUC of 0.97 (95% CI: 0.95–0.98) in the bootstrap resampling technique (n=1000) (Figure 3a). Supplementary Table 2 presents the general characteristics and distribution of included features in the test set patients. The model maintained its robust performance in the validation set with an AUC of 0.94 (95% CI: 0.86–1.00) (Figure 3b). The calibration plot demonstrated that the predictions closely matched the observed IBD probabilities, a finding further supported by the Hosmer–Lemeshow test results (P=0.19; Figure 3c). The model demonstrated good consistency in the external validation set, with the Hosmer–Lemeshow test results (P = 0.13) indicating a satisfactory fit, a well-calibrated curve, a calibration slope of 0.86, and a Brier score of 0.08 (Figure 3d). Decision curve analysis showed that the nomogram provided a higher net benefit than the IBD-REFER rule across a wide range of threshold probabilities (10–80%) in both the training and validation cohorts, indicating superior clinical utility for identifying children at risk of IBD (Supplementary Figure 3).

Figure 3 Evaluation of model stability and diagnostic performance in the training and validation datasets. (a) Receiver operating characteristic (ROC) curves of the nomogram model in the training dataset. (b) ROC curves of the nomogram model in the validation dataset. (c) Validation of the nomogram model for pediatric inflammatory bowel disease (IBD) in the training dataset. (d) Validation of the nomogram model for pediatric IBD in the validation dataset. Abbreviations: AUC, area under curve; CI, confidence interval.

When the cut-off value was set at 0.32, the model attained the highest Youden index, with a sensitivity of 91% and a specificity of 91%, significantly exceeding the performance of the IBD-REFER criteria (sensitivity, 99%; specificity, 34%). In validation datasets, the AUC was significantly higher in our model than in the IBD-REFER (AUC=0.94 [95% CI: 0.86–1] versus 0.67 [95% CI: 0.63–0.7]; P<0.01; Figure 3b).

In the sensitivity analysis, we explored the utility of the simplified versions of the model without the ultrasound features (Limberg level >1). The performance of the simplified versions was inferior to that of the models that incorporated the ultrasound features (AUC=0.94 [95% CI: 0.91–0.97] versus 0.97 [95% CI: 0.95–0.98]).

The predicted probabilities of being classified as IBD were found to have a moderate to strong positive association with clinical (ρ =0.569 and 0.463, P<0.01), endoscopic (ρ =0.631 and 0.603, P<0.01), and histological (ρ =0.693 and 0.66, P<0.01) severity of IBD in both the training and test sets (Supplementary Figure 4).

To assess the impact of single imputation, a complete-case sensitivity analysis was performed (n = 311). The final logistic regression model based on complete cases retained the same six predictors as the main model and yielded similar coefficients and intercept values (Supplementary Table 3). Both models achieved the same AUC of 0.94 (95% CI: 0.86–1) in the validation cohort.

Discussion

This study constructed a time- and cost-efficient digital model for the early identification of children with pediatric IBD on the basis of noninvasive test results. The developed model achieved good classification accuracy and showed robust predictive capacity through internal and temporal external validations. Additionally, we explored the relationship between the regression model and IBD clinical, endoscopic, and histological severity. To our best of knowledge, our study enrolled the largest sample size of pediatric IBD cases. The unique advantage of this study is that primary healthcare personnel can identify IBD at an early stage through noninvasive examinations, facilitating patient referral and assisting accelerated endoscopic diagnosis.

Delayed diagnosis is associated with growth failure, more extensive disease, poorer response to medical treatment, increased need for surgery, and decreased health-related quality of life.^27–29 Therefore, early identification and treatment of pediatric IBD are necessary. Unfortunately, the clinical manifestations of IBD are often nonspecific, and neither patient history nor physical examination can accurately distinguish IBD from functional or other gastrointestinal disorders.³⁰ The diagnostic delay in our cohort underscores this issue, and the access of children at high risk of pediatric IBD to specialist resources may be limited. Furthermore, the increasing global incidence of IBD pressurize healthcare systems to accommodate the growing number of patients with IBD.³¹ Developing and implementing tools to assist clinicians in identifying patients at high risk of IBD is a strategy that ensures timely diagnosis and rational allocation of healthcare resources.

The initial index is a questionnaire developed by the International Organization for IBD based on symptoms and signs alone, but it showed poor sensitivity and specificity when it was validated across centers.^10,32 Pediatricians also use noninvasive tools, including clinical disease activity indices and routine laboratory test results, such as CRP, ESR, and FC.³³ Although these tools serve as nonspecific markers of inflammation, their utility in isolation for diagnosis is limited. IBD-REFER has provided methods for distinguishing IBD on the basis of clinical symptoms and biochemical markers, aiding in early referral.¹⁰ However, in the training dataset of this study, the specificity of the IBD-REFER diagnosis for children appeared low. Compared with clinical symptoms and inflammatory markers, ultrasonography can provide more direct visualization of intestinal inflammatory changes. Its accuracy in diagnosing intestinal inflammation is limited if the lesions are superficial or confined to areas in which ultrasound detection is poor, eg, the small intestine or rectum.³⁴ Therefore, constructing a model with multiple levels of indicators is an appropriate option.

We included ultrasound indicators in the study model, and the importance ranking based on the random forest analysis indicated that bowel wall thickening with accompanying blood flow is the best diagnostic indicator for the classification models. Despite longer disease duration correlating with a higher probability of abnormalities in ultrasound, our study found that even in the early stages of disease (within 2 months), ultrasound can still achieve good diagnostic accuracy (Supplementary Figure 2). It is worth noting that, unlike clinical symptoms and biochemical markers, ultrasonography has weaknesses including operator dependency, which may also limit the generalizability of our model. However, a recent study³⁵ reported that pediatric gastroenterologists can achieve good consistency with ultrasound experts in detecting bowel wall thickening after 1 month of short-term training about intestinal ultrasonography, which undoubtedly demonstrates the accessibility and convenience of the ultrasound method. The inclusion of Limberg level >1, a transabdominal intestinal ultrasound indicator, in this model ensures the scanning of the entire intestine length rather than focusing on a single segment. This indicator explores the qualitative assessment of intestinal wall thickening with blood flow. Compared with other indicators such as intestinal wall layering and specific segmental intestinal wall thickness, the Limberg level requires lower operational demands, making it advantageous for implementation in primary healthcare settings.

There are some limitations in the current study. First, this was a retrospective single-center study, and no geographical external validation was conducted. Our hospital is China’s largest IBD diagnosis and treatment center, with many newly diagnosed and referred patients. The high prevalence of IBD among our study population may be beneficial for the diagnostic performance of the model, which limits the external validity of our diagnostic tool to an extent. Second, although the sample size of this study is larger than that of existing IBD-referred studies, the number of IBD diagnoses was still small compared with that of non-IBD diagnoses. Future research should be conducted in a larger, multi-center setting to validate and refine the model. Third, children differ from adults as there are distinct developmental stages during childhood. Owing to the limited sample size, we combined younger and older children in our discussion. However, future research should systematically explore early noninvasive diagnostic models to accommodate different age groups. Fourth, the applicability of our model may be limited in centers without access to experienced pediatric IUS operators. The use of intestinal ultrasound in pediatric care remains restricted in many settings due to the lack of trained personnel and standardized protocols, which may affect reproducibility and real-world implementation. Finally, incorporation bias may have occurred because IUS operators were not blinded to the patients’ clinical information. Although this reflects routine clinical practice, where sonographers often interpret ultrasound findings based on available clinical data, it may lead to an overestimation of diagnostic performance, particularly concerning the contribution of IUS-related variables to the model. Future studies should adopt blinded IUS assessments to more rigorously evaluate the model’s robustness and minimize the risk of observer bias.

Conclusions

We developed a model based on rapid, simple, and cost-effective noninvasive tests and demonstrated its potential as a digital tool for the early identification of children with IBD. This is particularly important for IBD management, as the disease can progress rapidly and lead to serious complications if not diagnosed and treated promptly. Our findings suggest that a Limberg score >1 plays a key role in distinguishing between IBD and non-IBD cases. While promising, the model requires external validation in multicenter studies across diverse clinical settings before it can be recommended for routine implementation. Importantly, it is designed to complement, not replace, endoscopic evaluation.