Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease

Abstract

Background:

Stricturing Crohn’s disease (CD) is heterogeneous, and risk factors for natural disease course progression remain unclear.

Objectives:

The study aims to construct a predictive model based on commonly used clinical indicators to identify high-risk stricturing CD prone to natural disease course progression.

Design:

Retrospective multicenter study.

Methods:

We conducted a retrospective analysis of clinical data from stricturing CD patients. The natural disease course progression was defined as progression to penetrating disease or surgery. The data were split into a test cohort and an internal validation cohort in a 7:3 ratio. Classification models, including logistic regression (LR), random forest (RF), and extreme gradient Boosting (XGB) models, were constructed and validated in an independent external validation cohort.

Results:

The study included 341 patients, with 190 in the internal training cohort, 81 in the test cohort, and 70 in the external validation cohort. Obstructive symptoms, globulin (GLB), erythrocyte sedimentation rate (ESR), and C-reactive protein (CRP) are independent predictors of natural disease course progression. LR, RF, and XGB models were developed to predict the occurrence of natural disease course progression within 2 years. In the internal test cohort, the areas under the receiver-operating characteristic curves of the LR, RF, and XGB models were 0.861, 0.492, and 0.559, respectively. In the external validation cohort, the LR model also achieved the highest area under the receiver operating characteristic curve value of 0.752.

Conclusion:

In patients with stricturing CD, obstructive symptoms score, CRP, ESR, and GLB were independent and validated predictors of natural disease course progression. A prognostic nomogram based on the LR model was developed to aid in evaluating the prognosis of stricturing CD patients.

Plain language summary

A simple tool to predict Crohn’s disease progression

Clinical records from patients with Crohn’s disease causing intestinal narrowing were reviewed to identify common signs that may indicate a higher risk of complications or the need for surgery. The analysis found that symptoms of intestinal blockage and higher levels of C-reactive protein, erythrocyte sedimentation rate, and globulin are linked to disease progression over a two-year period. A logistic regression model was used to create a practical tool for assessing future risk. This tool offers a straightforward method for healthcare providers to evaluate patient prognosis and inform treatment decisions.

Keywords

Crohn’s disease natural disease course progression prediction model stricture

Introduction

Crohn’s disease (CD) is a chronic inflammatory disorder of the gastrointestinal tract. Approximately 20% of CD patients initially present with a stricturing phenotype, and over 50% experience stricturing complications during the natural course of the disease.^1,2 Currently, treatment options for intestinal strictures include conventional pharmacotherapy, biologic agents such as Ustekinumab,³ and endoscopic balloon dilatation. Despite these interventions, strictures can progress to more severe disease states (penetrating disease) and even need surgery.⁴ A key contributing factor to this unfavorable trajectory is the absence of a unified and precise definition of stricturing disease, which hampers timely diagnosis and effective risk stratification. Recent consensus defined stricturing CD: the combination of luminal narrowing, pre-stricture dilation, and wall thickness based on computed tomography (CT), magnetic resonance imaging, or intestinal ultrasonography, and the inability to pass an adult or pediatric colonoscope with a reasonable amount of pressure.³ Previous studies on CD-related strictures based on consensus also mostly included patients with proximal intestinal dilation, which is a severe condition and associated with a high risk of natural disease course progression.⁵ This limited our understanding of the natural history of CD-related strictures.^6–8 Patients with strictures but without proximal bowel dilation also deserve close monitoring, as they may represent an earlier stage of disease evolution that warrants proactive intervention. Including early-stage CD patients with intestinal strictures could provide valuable insights into the full course of stricture development in CD.

Elevated serum markers such as erythrocyte sedimentation rate (ESR), platelet count (PLT), C-reactive protein (CRP), and globulin (GLB) have been associated with stenosis.^9–11 Furthermore, increased CRP levels and serological markers such as perinuclear antineutrophil cytoplasmic and anti-Saccharomyces cerevisiae antibodies have predictive values for stricture development.^10,12 Symptoms, including acute abdominal pain, have also been identified as a significant marker of poor prognosis in CD-associated strictures. In a recent study by El Ouali et al.,¹³ multivariable analysis highlighted that the intestinal obstruction symptom score, including acute abdominal pain and bloating, was an independent and validated predictor of the need for intervention. While emerging omics such as genomics and proteomics approaches show promise for evaluating the risk of stricture, their complexity and high cost currently limit their application in routine clinical settings.¹⁴ Thus, there is an urgent need to identify accessible and reliable predictive factors to develop clinically applicable predictive models for the early detection of stricturing CD patients with poor prognosis, particularly those without proximal bowel dilation.

In recent years, machine learning (ML) models involving clinical data from endoscopy and imaging have shown potential in predicting natural disease course progression in CD.^5,15 Building on this progress, this study aims to identify risk factors for natural disease course progression, such as penetration or surgery in stricturing CD patients, and to develop a simple and practical prediction model for early intervention of these patients.

Materials and methods

Figure 1 illustrates the overall study design.

Figure 1.

Overall study flowchart.

Participants

This is a dual-center retrospective cohort study. The internal cohort was obtained from the clinical database of the First Affiliated Hospital, Sun Yat-sen University (FAH-SYSU), collected by retrieving patients’ outpatient, physical examination, and hospitalization records. The diagnosis of CD was established according to the European Crohn’s and Colitis Organisation guidelines.¹⁶ The inclusion criteria for the study are as follows: (1) age ⩾18 and (2) with intestinal strictures confirmed by clinical symptoms, imaging, or endoscopic evaluation at each center. Intestinal stricture was defined by the presence of any of the following criteria: (2.1) Persistent luminal stricture with inability to pass a standard-sized endoscope and (2.2) imaging (MR/CT) showing intestinal stricture, with or without proximal bowel dilation.¹⁷ (2.3) Significant symptoms of intestinal obstruction include vomiting, abdominal pain, abdominal distension, and cessation of gas and stool passage.

The exclusion criteria for the study are as follows: (1) with stricturing CD complicated by rectal tumors or infectious colitis; (2) progression to stricturing CD within 3 months from initial diagnosis; and (3) a history of intestinal resection. An independent external validation cohort of the same time frame was recruited from Shanghai Jiao Tong University School of Medicine (Ruijin Hospital). A uniform set of criteria was used to build the two cohorts. Research approval was obtained from the Ethics Committees of the FAH-SYSU (approval no. 2023-488).

Data collection and definitions

Based on previous studies and the characteristics of the clinical data at our center, this study retrospectively collected 31 clinical indicators. Due to missing data for some indicators, 24 clinical indicators were ultimately included for analysis. These include the following: basic patient information and disease-related data (gender, L classification in the Montreal classification, age at CD diagnosis), clinical symptoms (abdominal pain, postprandial abdominal pain, acute abdominal distension, abdominal distension, abdominal cramps, dietary restrictions, nausea, and/or vomiting), medication history and personal history (history of steroid, immunosuppressant or biologic use, smoking history), and laboratory indicators (white blood cell (WBC), PLT, hemoglobin (HB), alkaline phosphatase (ALP), GLB, alanine aminotransferase, aspartate transaminase (AST), albumin (ALB), ESR, CRP). Laboratory indicators were obtained from the pre-intervention test results (window ±7 days). The classification of clinical symptoms was based on an obstructive symptom scoring questionnaire applicable to CD patients. Based on the previous study by El Ouali et al.,¹³ we combined the seven collected clinical symptoms into a composite score for obstructive symptoms (including abdominal pain, postprandial abdominal pain, acute abdominal distension, bloating, abdominal cramping, dietary restrictions, nausea, and/or vomiting), ranging from 0 to 7. A smoker was defined as someone who actively smoked at least 7 cigarettes/week.¹⁸

Definition of baseline and outcome

The first report of intestinal stricture was defined as the baseline. The natural disease course progression was defined as progression to penetration or surgery due to intestinal strictures. Follow-up duration was defined as the time from the endoscopic/imaging report (MR/CT) of stricture to surgery, the first clinical diagnosis of penetration, loss to follow-up, or the end of the study (October 30, 2023).

Sample size estimation

Sample size estimation was performed using the R package “pmsampsize.” With an anticipated C-index of 0.9 and a restriction to no more than 10 predictors for model simplification, and based on previous studies indicating a positive event rate of 15%–20%,^19–21 the required sample size was determined to be between 255 and 293 subjects. The study cohort was subsequently divided into a training cohort and an independent validation cohort in a 7:3 ratio.

Construction of logistic, random forest, and extreme gradient boosting models

To establish a risk prediction model for the progression of stricturing CD, the FAH-SYSU cohort were randomly divided into a training cohort and an internal validation cohort at a ratio of 7:3. The proportion of missing values across all baseline variables was less than 5%, and multiple imputation was performed on the missing data using the R package “mice.” A comprehensive summary of missing data per variable and per participant is available in Supplemental Table 1. The Ruijin Hospital cohort was considered an independent validation cohort. All models were developed and validated using R 4.3.3 (RStudio, Boston, MA, USA). The study followed the TRIPOD guidelines to ensure standardized reporting of prediction model development and validation using logistic regression (LR) and ML approaches.²²

Univariate LR was applied to establish LR models for each independent variable and the outcome of interest. Significant statistical variables in univariate analyses were subsequently included as independent variables in a multivariate LR model. Variables for multivariate analysis were selected based on clinical relevance and biological plausibility.²³

The random forest (RF) was a bagging ensemble ML method based on binary recursive decision trees. For the RF model, the selected hyperparameters were ntrees = 200 and mtry = 15. These values were similarly determined based on training set performance. Variable importance was measured using the Mean Decrease Gini. The R packages “randomforest” and “caret” were used to develop and validate the RF model.

Extreme gradient boosting (XGB) was an ensemble learning model that used a gradient boosting framework based on decision trees. For the XGB model, we performed grid search over a predefined range of values and selected the final hyperparameters based on training performance: learning_rate = 0.08, max_depth = 6, max_delta_step = 8, min_child_weight = 3, nround = 1, subsample = 0.68, lambda = 0.5, colsample_bytree = 0.75, colsample_bylevel = 0.75, colsample_bynode = 0.75, and scale_pos_weight = 9 (to account for class imbalance). The R package “xgboost” was used for the development and validation of the XGB model. In the XGB model, hyperparameter tuning was performed. Variable importance was measured using the Gain.

Evaluation and validation of the models

The area under the receiver operating characteristic curve (AUC) was used to evaluate the primary performance of the model. The DeLong test was performed to compare the AUC of different models. Finally, the best model (AUCs were greater than 0.75 in all different sets) was selected and converted into a nomogram to enhance its clinical practical value. Decision curve analysis (DCA) was performed to assess the clinical net benefit of the model by considering both true-positive and false-positive rates across different threshold probabilities. In addition, calibration curves were constructed to evaluate the agreement between predicted and observed outcomes, providing insights into the accuracy of the model.

Construction of survival curves

The predicted probabilities from the LR model were stratified using the “surv_cutpoint” function from the R package “survminer.” Patients in the nomogram risk stratification group with a risk score greater than the cutoff value were classified as high-risk, while others were classified as low risk. Survival curves were constructed using the Kaplan–Meier method to estimate the survival function for different groups. The log-rank test was employed to compare survival differences between groups.

Statistical analysis

Data description and analysis were performed using R 4.3.3. Continuous variables with a non-Gaussian distribution are expressed as the median and interquartile range (IQR), and group differences were assessed using the Mann–Whitney U test. Categorical variables were expressed as percentages, and group differences were determined using Fisher’s exact test. A p-value of <0.05 was considered to indicate statistical significance.

Results

Patient demographics and disease characteristics

A total of 580 patients with stricturing CD were initially assessed, out of which 416 patients met the inclusion criteria and were enrolled in the study cohort. Subsequently, 142 patients were excluded based on the exclusion criteria, including (1) structuring CD complicated by rectal tumors or infectious colitis (n = 4), (2) progression to structuring disease within 3 months of initial diagnosis (n = 83), and (3) a history of intestinal resection (n = 55). Three patients were further excluded due to incomplete electronic medical records, leaving a total of 271 patients for data analysis. Figure 2 shows the flowchart of patient inclusion and exclusion. The median follow-up period was 22 months. During our follow-up period, 27 patients (10%) experienced disease progression to penetrating disease or surgery intervention. Characteristics of baseline patients in the internal cohort and the external cohort are shown in Tables 1 and 2, respectively. Progress rates for priority events are shown in Supplemental Table 2.

Figure 2.

Flowchart of patient inclusion and exclusion.

Table 1.

Baseline characteristics in the internal cohort.

Characteristics	Internal cohort
	Internal cohort overall	Non-progression	Progression	p-Value^a
	(N = 271)	(N = 244)	(N = 27)	p-Value^a
Location
L1	36 (13.3%)	32 (13.1%)	4 (14.8%)	0.61
L2	18 (6.6%)	18 (7.4%)	0 (0%)
L3	166 (61.3%)	148 (60.7%)	18 (66.7%)
L4	51 (18.8%)	46 (18.9%)	5 (18.5%)
Sex
Male	190 (70.1%)	172 (70.5%)	18 (66.7%)	0.663
Female	81 (29.9%)	72 (29.5%)	9 (33.3%)
Age at diagnosis
Median (IQR)	26 (20, 34)	26 (20, 34)	27 (24, 34)	0.268
Steroid use
No	212 (78.2%)	189 (77.5%)	23 (85.2%)	0.465
Yes	59 (21.8%)	55 (22.5%)	4 (14.8%)
Immunosuppressants use
No	183 (67.5%)	163 (66.8%)	20 (74.1%)	0.521
Yes	88 (32.5%)	81 (33.2%)	7 (25.9%)
Biologics use
No	202 (74.5%)	182 (74.6%)	20 (74.1%)	0.581
IFX	56 (20.7%)	51 (20.9%)	5 (18.5%)
ADA	8 (3.0%)	6 (2.5%)	2 (7.4%)
VDZ	2 (0.7%)	2 (0.8%)	0 (0%)
UST	3 (1.1%)	3 (1.2%)	0 (0%)
Smoking
No	242 (89.3%)	219 (89.8%)	23 (85.2%)	0.508
Yes	29 (10.7%)	25 (10.2%)	4 (14.8%)
WBC
Median (IQR)	6.53 (5.18, 7.98)	6.64 (5.24, 8.05)	6.10 (4.85, 7.72)	0.411
PLT
Median (IQR)	323 (258, 394)	318 (254, 390)	351 (301, 426)	0.082
HB
Median (IQR)	117 (101, 132)	119 (101, 132)	114 (99, 128)	0.419
ALP
Median (IQR)	71.0 (62.0, 83.5)	70.5 (61.8, 82.0)	77.0 (64.0, 86.5)	0.368
GLB
Median (IQR)	29.6 (25.6, 33.5)	29.3 (25.4, 33.0)	31.8 (29.0, 36.0)	0.009
ALT
Median (IQR)	11.0 (7.5, 16.0)	11.0 (8.0, 16.0)	9.00 (6.5, 12.5)	0.157
AST
Median (IQR)	16.0 (13.0, 20.0)	16.0 (13.0, 20.0)	14.0 (12.0, 17.5)	0.017
ALB
Median (IQR)	35.1 (32.35, 38.7)	35.7 (32.6, 39.0)	33.5 (31.75, 36.6)	0.056
ESR
Median (IQR)	31.0 (15.0, 54.0)	29.0 (14.0, 53.3)	50.0 (26.0, 62.0)	0.025
CRP
Median (IQR)	14.3 (3.11, 33.1)	13.7 (3.0, 32.4)	21.3 (8.9, 45.3)	0.049
Obstructive symptoms score
Median (IQR)	1 (1, 2)	1.00 (1.00, 2.00)	2 (1, 2)	<0.001

p-Value: Differences between revert to B3 or surgery and non-progression groups in the internal cohort.

ADA, adalimumab; ALB, albumin; ALP, alkaline phosphatase; ALT, alanine aminotransferase; AST, aspartate transaminase; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; GLB, globulin; HB, hemoglobin; IFX, infliximab; IQR, interquartile range; PLT, platelet count; UST, ustekinumab; VDZ, vedolizumab; WBC, White blood cell.

Table 2.

Baseline characteristics in the external cohort.

Characteristics	External cohort
	External cohort overall	Non-progression	Progression	p-Value^a
	(N = 70)	(N = 27)	(N = 43)	p-Value^a
Location
L1	20 (28.6%)	8 (29.6%)	12 (27.9%)	0.343
L2	5 (7.1%)	2 (7.4%)	3 (7.0%)
L3	24 (34.3%)	6 (22.2%)	18 (41.9%)
L4	21 (30.0%)	11 (40.7%)	10 (23.3%)
Sex
Male	55 (78.6%)	24 (88.9%)	31 (72.1%)	0.137
Female	15 (21.4%)	3 (11.1%)	12 (27.9%)
Age at diagnosis
Median (IQR)	28 (25, 39)	26 (22, 32)	31 (26, 42)	0.028
Steroid use
No	52 (74.3%)	22 (81.5%)	30 (69.8%)	0.401
Yes	18 (25.7%)	5 (18.5%)	13 (30.2%)
Immunosuppressants use
No	50 (71.4%)	20 (74.1%)	30 (69.8%)	0.79
Yes	20 (28.6%)	7 (25.9%)	13 (30.2%)
Biologics use
No	20 (28.6%)	10 (37.0%)	10 (23.3%)	0.019
IFX	26 (37.1%)	14 (51.9%)	12 (27.9%)
ADA	16 (22.9%)	2 (7.4%)	14 (32.6%)
VDZ	1 (1.4%)	0 (0%)	1 (2.3%)
UST	7 (10.0%)	1 (3.7%)	6 (14.0%)
Smoking
No	62 (88.6%)	23 (85.2%)	39 (90.7%)	0.702
Yes	8 (11.4%)	4 (14.8%)	4 (9.3%)
WBC
Median (IQR)	5.74 (4.66, 7.89)	6.71 (5.24, 8.89)	5.63 (4.43, 7.02)	0.114
PLT
Median (IQR)	296 (229, 359)	300 (260, 352)	296 (227, 362)	0.682
HB
Median (IQR)	118 (100, 131)	118 (93, 131)	119 (102, 130)	0.616
ALP
Median (IQR)	67.5 (61.0, 84.8)	66 (55, 78)	70 (62, 92)	0.101
GLB
Median (IQR)	30.5 (27.0, 35.0)	30 (26, 35)	31 (27, 36)	0.608
ALT
Median (IQR)	9.50 (7.0, 16.0)	8 (7, 14)	13 (7, 20)	0.284
AST
Median (IQR)	16.5 (12.2, 21)	15 (12, 18)	18 (13, 22)	0.174
ALB
Median (IQR)	36.0 (32.2, 40.0)	37 (34, 40)	36 (31, 40)	0.131
ESR
Median (IQR)	11.0 (7.0, 20.0)	12 (6, 18)	10 (7, 23)	0.595
CRP
Median (IQR)	11.9 (3.6, 56.1)	11.0 (2.5, 54.2)	15.3 (4.1, 55.2)	0.375
Obstructive symptoms score
Median (IQR)	2 (1, 3)	1 (1, 2)	3 (2, 4)	<0.001

p-Value: Differences between revert to B3 or surgery and non-progression groups in the external cohort.

ALB, albumin; ALP, alkaline phosphatase; ALT, alanine aminotransferase; AST, aspartate transaminase; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; GLB, globulin; HB, hemoglobin; IQR, interquartile range; PLT, platelet count; WBC, White blood cell.

Model development and validation

A total of 271 patients were collected from FAH-SYSU, and 70 patients from Ruijin Hospital were included in the final analysis. There were no significant differences in baseline characteristics between the training cohort and the internal validation cohorts. Comparison of variable characteristics between the two cohorts is presented in Supplemental Table 3. In the internal development and validation cohorts, there were no significant differences in baseline clinical variables, supporting the integrity of the randomization and comparability of disease severity (Supplemental Table 3). Differences observed between the internal and external validation cohorts reflect real-world heterogeneity between centers and emphasize the generalizability of the proposed model across diverse clinical settings (Supplemental Table 3). The generalization ability of the models was evaluated using internal and external validation cohorts.

Logistic regression

After performing univariate LR analysis, variables with a p-value less than 0.05 (obstructive symptoms score, GLB, ESR, and CRP) were selected for inclusion in a multivariate LR model to assess their association with the outcome (Table 3). The final optimized model achieved the following: AUC = 0.768 (95% confidence interval (CI): 0.659–0.877) for the training set, AUC = 0.861 (95% CI: 0.753–0.969) for the internal validation, and AUC = 0.752 (95% CI: 0.637–0.867) for the external validation (Figure 3). In the multivariate LR model, obstructive symptoms score was identified as the only independent risk factor (Table 3).

Table 3.

Univariate and multivariate logistic regression analyses.

Variable	Univariate analysis				Multivariate analysis
Variable	Beta	Lower 95% CI	Upper 95% CI	p Value	Beta	p Value
Location	−15.96	NA	55.96	0.989	/	/
Sex	0.31	−0.72	1.27	0.529	/	/
Age at diagnosis	0.02	−0.03	0.05	0.462	/	/
Steroid use	−0.56	−2.05	0.60	0.394	/	/
Immunosuppressants use	−0.57	−1.73	0.43	0.292	/	/
Biologics use	−0.04	−1.35	1.05	0.942	/	/
Smoking	−7.87E−16	−1.89	1.36	1.000	/	/
WBC	0.13	−0.07	0.31	0.189	/	/
PLT	3.63E−03	0.00	0.01	0.075	/	/
HB	−0.01	−0.03	0.01	0.277	/	/
ALP	−2.42E−03	−0.03	0.01	0.805	/	/
GLB	0.10	0.02	0.17	0.012	0.07	0.124
ALT	−0.03	−0.09	0.02	0.352	/	/
AST	−0.08	−0.18	2.75E−03	0.095	/	/
ALB	−0.09	−0.18	3.16E−03	0.060	/	/
ESR	0.02	9.37E−04	0.03	0.035	0.01	0.577
CRP	0.02	7.38E−04	0.03	0.033	0.01	0.595
Obstructive symptoms score	0.83	0.38	1.35	0.001	0.91	0.001

ALB, albumin; ALP, alkaline phosphatase; ALT, alanine aminotransferase; AST, aspartate transaminase; CI, confidence interval; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; GLB, globulin; HB, hemoglobin; PLT, platelet count; WBC, White blood cell.

Figure 3.

Performance of the logistic regression model in the internal training set, internal validation set, and external independent validation set.

Random forest

The RF model was constructed using all 18 independent variables ultimately included in the study. In the training set, RF achieved an AUC 0.908 (95% CI = 0.835–0.981). The internal validation yielded an AUC of 0.492 (95% CI = 0.313–0.672), while the external validation resulted in an AUC of 0.647 (95% CI = 0.518–0.776; Supplemental Figure 1(A)). According to the Mean Decrease Gini, the five most predictive variables identified by the RF were obstructive symptoms score, CRP, WBC, GLB, AST, and age at diagnosis. The top important variables were displayed in Supplemental Figure 1(B).

Extreme gradient boosting

The XGB model based on the full variables achieved an AUC of 0.785 (95% CI: 0.690–0.881) for the training set, an AUC of 0.559 (95% CI: 0.354–0.764) for the internal validation cohort, and an AUC of 0.608 (95% CI: 0.492–0.723) for the external validation cohort (Supplemental Figure 2(A)). The six most predictive variables identified were GLB, ALP, obstructive symptoms score, ESR, sex, and ALB (Supplemental Figure 2(B)).

Model comparison

A comparison of the LR, RF, and XGB models revealed that the LR model achieved the highest predictive performance, with an AUC of 0.861 in internal validation and 0.752 in external validation. The DeLong test confirmed that the AUC of the LR model was statistically significantly higher than those of the RF and XGB models, with p-values of 6.065e−06 and 0.012, respectively.

Nomogram conversion and patient stratification

Considering the clinical applicability of the model, we performed a Nomogram conversion of the optimal LR model to assist clinicians in their decision-making. Detailed performance of the LR model is provided in Supplemental Table 4. The converted nomogram is shown in Figure 4. DCA and calibration curves were conducted using the nomogram for both the internal and external validation cohorts, and results demonstrated the robustness of the nomogram (Supplemental Figures 3 and 4). In addition, patients predicted by the Nomogram to have a greater than cutoff were defined as high risk, while the remaining patients were classified as low risk. Survival curve of the external cohort analysis indicated the nomogram could accurately stratify patients at risk of disease progression (p = 0.00021; Supplemental Figure 5). Furthermore, revalidation in the entire cohort demonstrated that the nomogram could robustly stratify patients based on their risk of natural disease course progression (p < 0.0001; Figure 5).

Figure 4.

Nomogram depicting the probability of natural disease course progression.

Figure 5.

Survival curve analysis with nomogram classification of risk.

Discussion

Approximately 50% of CD patients were at risk of developing strictures over their lifetime. By constructing a model to accurately identify patients with rapid progression in stricturing CD, personalized diagnosis and treatment can be achieved, reversing the natural disease course and improving the quality of life for these patients. In this study, we included patients with early-stage CD strictures and constructed a predictive model based on readily available clinical indicators, accurately forecasting the risk of disease progression in CD strictures. The prediction tools can assist clinicians in the early identification of stricturing CD patients who are at risk of rapid progression, thereby improving prognosis and potentially reversing the natural course of the disease.

In this study, candidate variables for the model were selected based on existing evidence and clinical relevance from baseline study visits. A risk prediction model was constructed using an LR model, a synthetic minority oversampling technique-based RF model, and an XGB model. After comparing the three models, the logistic model performed the best in both internal and external validations, achieving the highest AUC (internal validation AUC = 0.861, external validation AUC = 0.752). Consequently, the LR model was adopted as the risk prediction model for the progression of stricturing CD. The results from the LR were converted into a nomogram to aid clinical decision-making.²⁴ In addition, patients with a nomogram-predicted probability of disease progression greater than the cutoff were defined as high risk, while others were classified as low risk. Survival curve analysis demonstrated the excellent discriminatory ability of the nomogram in patient stratification.

Clinical symptoms such as abdominal cramping, nausea, or vomiting have been identified in previous studies as predictors for surgical or endoscopic balloon dilation interventions in patients with ileal stricturing CD. In the study by El Ouali et al.,¹³ the obstruction score (hazard ratio (HR) 1.40 (1.10–1.79); p = 0.01) was considered a potentially useful tool that is easy to use in practice and assists in clinical decision-making. We referred to this study and used the obstructive symptoms score as a predictor. In our study, clinical symptoms for obstruction were scored as the most important predictor of stricturing CD progression, being the most significant variable in both the LR and RF models. The intestinal stricture obstruction score was also identified as the only independent risk factor in the multivariate LR model.

Serum inflammatory biomarkers have been reported to be associated with the stricturing phenotype. Li et al.¹⁰ found that elevated serum ESR could predict both intestinal strictures and fistulas in CD. Henriksen et al.²⁵ discovered a significant association between CRP levels at diagnosis and surgical risk in patients with CD and terminal ileitis, with elevated CRP levels above 53 mg/L increasing the risk (odds ratio = 6.0, 95% CI (1.1–31.9), p = 0.03). Previous study also found that elevated GLB levels are important predictors of increased hospitalization rates and IBD-related surgeries in CD patients.¹¹ In our study, we focused on the predictive value of ESR, CRP, and GLB levels in the rapid progression of intestinal strictures in CD, and achieved good predictive results.

ML techniques have been successfully applied to various topics related to IBD, including identifying risk genes, predicting outcomes from serum proteins, and predicting drug response from multi-omics data.^26–28 However, they have not been specifically applied to studies on disease progression within the fixed intestinal stricture phenotype. El Ouali et al.¹³ used multivariate analysis to identify risk factors for intervention within 1 and 4 years in CD, including stricture duration, length, and obstruction index, and constructed a risk prediction model based on these factors. In comparison, our study focused specifically on the intestinal stricture phenotype in CD, constructing and comparing three prediction models, ultimately selecting the best-performing model. This approach makes the results more precise.

Furthermore, a systematic review highlighted several limitations in studies applying ML techniques to develop clinical outcome prediction models in IBD patients. While all studies attempted to address potential model overfitting through internal validation, only one study attempted external validation by relying on an external patient cohort.^29,30 Our study employed both internal and external validation, which enhances the credibility of the findings. Notably, the performance of the traditional LR model in our study surpassed that of other ML models, which contradicts some previous findings.²⁹ We offer the following explanation for these discrepancies: ML models often encounter class imbalance issues during development, which clinical settings are not well-suited to adjust, potentially leading to distortions in disease prevalence.³¹ In addition, many studies have shown that ML models require more data than traditional LR models to perform well.³² The superior performance of the LR model in our study does not negate the usability and potential of ML-based models in prediction. As clinical data rapidly grow and multi-omics high-dimensional data models evolve, ML models still hold great promise for future applications in medical modeling.³³

A nomogram incorporating four parameters was developed to accurately predict natural disease course progression in CD patients with intestinal strictures. In clinical implementation, several considerations must be addressed. First, rigorous evaluation of input data quality is essential. Predictor values that are poor in quality or missing should be identified using predefined metrics and managed appropriately via imputation techniques or exclusion criteria. Second, data entry is performed by clinicians based on actual clinical conditions, ensuring that the model is seamlessly integrated into routine practice without demanding extensive technical expertise. Finally, future investigations should prioritize external validation in diverse clinical settings to further establish the applicability and generalizability.

Nonetheless, this study has several limitations. First, as a retrospective study, it relies on historical records, which may be incomplete or inaccurate, leading to selection and information biases. Prospective studies are needed for further validation. Second, the clinical decision prediction factors ultimately selected in this study were derived from the LR model, which inherently has some common flaws, such as overfitting to noise and outliers in the training data. To minimize the risk of overfitting, we selected another independent center for validation. In the future, the LR model could be integrated into ensemble learning methods and combined with other algorithms to improve overall predictive performance. Finally, the variables collected in this study did not include genetic or radiomics data, limiting the ability to identify and predict rapid progression in patients with stricturing CD at the genetic and imaging levels. Future studies could explore these aspects further.

In conclusion, this study included patients with early-stage CD strictures and developed an LR model to predict the progression risk of stricturing CD using clinical and serological indicators. Obstructive symptom scores, CRP, ESR, and GLB were identified as independent and validated predictors of natural disease course progression in stricturing CD. These findings provide a valuable tool for the early identification of patients at risk of rapid disease progression, enabling timely prevention and intervention.

Supplemental Material

sj-docx-1-tag-10.1177_17562848251358705 – Supplemental material for Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease

Supplemental material, sj-docx-1-tag-10.1177_17562848251358705 for Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease by Yu Wang, Xiaomin Wu, Weitong Gao, Xiaolong Chen, Haoyin Chen, Zishan Liu, Yao Zhang, Yubei Gu and Ren Mao in Therapeutic Advances in Gastroenterology

Supplemental Material

sj-tif-2-tag-10.1177_17562848251358705 – Supplemental material for Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease

Supplemental material, sj-tif-2-tag-10.1177_17562848251358705 for Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease by Yu Wang, Xiaomin Wu, Weitong Gao, Xiaolong Chen, Haoyin Chen, Zishan Liu, Yao Zhang, Yubei Gu and Ren Mao in Therapeutic Advances in Gastroenterology

Supplemental Material

sj-tif-3-tag-10.1177_17562848251358705 – Supplemental material for Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease

Supplemental material, sj-tif-3-tag-10.1177_17562848251358705 for Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease by Yu Wang, Xiaomin Wu, Weitong Gao, Xiaolong Chen, Haoyin Chen, Zishan Liu, Yao Zhang, Yubei Gu and Ren Mao in Therapeutic Advances in Gastroenterology

Supplemental Material

sj-tif-4-tag-10.1177_17562848251358705 – Supplemental material for Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease

Supplemental material, sj-tif-4-tag-10.1177_17562848251358705 for Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease by Yu Wang, Xiaomin Wu, Weitong Gao, Xiaolong Chen, Haoyin Chen, Zishan Liu, Yao Zhang, Yubei Gu and Ren Mao in Therapeutic Advances in Gastroenterology

Supplemental Material

sj-tif-5-tag-10.1177_17562848251358705 – Supplemental material for Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease

Supplemental material, sj-tif-5-tag-10.1177_17562848251358705 for Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease by Yu Wang, Xiaomin Wu, Weitong Gao, Xiaolong Chen, Haoyin Chen, Zishan Liu, Yao Zhang, Yubei Gu and Ren Mao in Therapeutic Advances in Gastroenterology

Supplemental Material

sj-tif-6-tag-10.1177_17562848251358705 – Supplemental material for Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease

Supplemental material, sj-tif-6-tag-10.1177_17562848251358705 for Development and validation of a novel model based on clinical characteristics to predict natural disease course progression in patients with stricturing Crohn’s disease by Yu Wang, Xiaomin Wu, Weitong Gao, Xiaolong Chen, Haoyin Chen, Zishan Liu, Yao Zhang, Yubei Gu and Ren Mao in Therapeutic Advances in Gastroenterology

Footnotes

Acknowledgements

Figures were created with .

Declarations

ORCID iD

Yu Wang

Supplemental material

Supplemental material for this article is available online.

References

Thia

Sandborn

Harmsen

, et al. Risk factors associated with progression to intestinal complications of Crohn’s disease in a population-based cohort. Gastroenterology 2010; 139(4): 1147–1155.

Yoo

Holubar

Rieder

. Fibrostenotic strictures in Crohn’s disease. Intest Res 2020; 18(4): 379–401.

Bettenworth

Baker

Fletcher

, et al. A global consensus on the definitions, diagnosis and management of fibrostenosing small bowel Crohn’s disease in clinical practice. Nat Rev Gastroenterol Hepatol 2024; 21(8): 572–584.

Revés

Mascarenhas

José Temido

, et al. Early intervention with biologic therapy in Crohn’s disease: how early is early? J Crohns Colitis 2023; 17(11): 1752–1760.

Inoue

Bartlett

Shahraki

, et al. Predicting risk of surgery in patients with small bowel Crohn’s disease strictures using computed tomography and magnetic resonance enterography. Inflamm Bowel Dis 2022; 28(11): 1677–1686.

Bouhnik

Carbonnel

Laharie

, et al. Efficacy of adalimumab in patients with Crohn’s disease and symptomatic small bowel stricture: a multicentre, prospective, observational cohort (CREOLE) study. Gut 2018; 67(1): 53–60.

Allocca

Bonifacio

Fiorino

, et al. Efficacy of tumour necrosis factor antagonists in stricturing Crohn’s disease: a tertiary center real-life experience. Dig Liver Dis 2017; 49(8): 872–877.

Pelletier

Kalisazan

Wienckiewicz

, et al. Infliximab treatment for symptomatic Crohn’s disease strictures. Aliment Pharmacol Ther 2009; 29(3): 279–285.

Lewis

. The utility of biomarkers in the diagnosis and therapy of inflammatory bowel disease. Gastroenterology 2011; 140(6): 1817–1826.e2.

10.

Qian

Bai

, et al. Prediction of complications in inflammatory bowel disease using routine blood parameters at diagnosis. Ann Transl Med 2022;10(4): 185.

11.

Hashash

Koutroumpakis

Anderson

, et al. Elevated serum globulin fraction as a biomarker of multiyear disease severity in inflammatory bowel disease. Ann Gastroenterol 2022; 35(6): 609–617.

12.

Reese

Constantinides

Simillis

, et al. Diagnostic precision of anti-Saccharomyces cerevisiae antibodies and perinuclear antineutrophil cytoplasmic antibodies in inflammatory bowel disease. Am J Gastroenterol 2006; 101(10): 2410–2422.

13.

El Ouali

Baker

Lyu

, et al. Validation of stricture length, duration and obstructive symptoms as predictors for intervention in ileal stricturing Crohn’s disease. United European Gastroenterol J 2022; 10(9): 958–972.

14.

Fiocchi

. Omics and multi-omics in IBD: no integration, no breakthroughs. Int J Mol Sci 2023; 24(19): 14912.

15.

Shi

Wang

Y di

Shen

X di

, et al. Clinical outcome is distinct between radiological stricture and endoscopic stricture in ileal Crohn’s disease. Eur Radiol 2023; 33(11): 7595–7608.

16.

Gomollón

Dignass

Annese

, et al. 3rd European evidence-based consensus on the diagnosis and management of Crohn’s disease 2016: Part 1: Diagnosis and medical management. J Crohns Colitis 2017; 11(1): 3–25.

17.

Gordon

Bettenworth

Bokemeyer

, et al. International consensus to standardise histopathological scoring for small bowel strictures in Crohn’s disease. Gut 2022; 71(3): 479–486.

18.

Louis

Michel

Hugot

, et al. Early development of stricturing or penetrating pattern in Crohn’s disease is influenced by disease location, number of flares, and smoking but not by NOD2/CARD15 genotype. Gut 2003; 52(4): 552–557.

19.

Vernier-Massouille

Balde

Salleron

, et al. Natural history of pediatric Crohn’s disease: a population-based cohort study. Gastroenterology 2008; 135(4): 1106–1113.

20.

Burisch

Kiudelis

Kupcinskas

, et al. Natural disease course of Crohn’s disease during the first 5 years after diagnosis in a European population-based inception cohort: an Epi-IBD study. Gut 2019; 68(3): 423–433.

21.

Gonczi

Lakatos

Kurti

, et al. Incidence, prevalence, disease course, and treatment strategy of Crohn’s disease patients from the Veszprem Cohort, Western Hungary: a population-based inception cohort study between 2007 and 2018. J Crohns Colitis 2023; 17(2): 240–248.

22.

Collins

Moons

KGM

Dhiman

, et al. TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 2024; 385: e078378.

23.

Yaw

Burrell

. Long-term outcomes and determinants of new-onset mental health conditions after trauma. JAMA Netw Open 2025; 8(3): e250349.

24.

Park

. Nomogram: an analogue tool to deliver digital knowledge. J Thorac Cardiovasc Surg 2018; 155(4): 1793.

25.

Henriksen

Jahnsen

Lygren

, et al. C-reactive protein: a predictive factor and marker of inflammation in inflammatory bowel disease. Results from a prospective population-based study. Gut 2008; 57(11): 1518–1523.

26.

Isakov

Dotan

Ben-Shachar

. Machine learning-based gene prioritization identifies novel candidate risk genes for inflammatory bowel disease. Inflamm Bowel Dis 2017; 23(9): 1516–1523.

27.

Ungaro

, et al. Machine learning identifies novel blood protein predictors of penetrating and stricturing complications in newly diagnosed paediatric Crohn’s disease. Aliment Pharmacol Ther 2021; 53(2): 281–290.

28.

Gardiner

Carrieri

Bingham

, et al. Combining explainable machine learning, demographic and multi-omic data to inform precision medicine strategies for inflammatory bowel disease. PLoS One 2022; 17(2): e0263248.

29.

Nguyen

Picetti

Dulai

, et al. Machine learning-based prediction models for diagnosis and prognosis in inflammatory bowel diseases: a systematic review. J Crohns Colitis 2022; 16(3): 398–413.

30.

Dias

Rodrigues

Coelho

, et al. Erratum: development and validation of risk matrices for Crohn’s disease outcomes in patients who underwent early therapeutic interventions. J Crohns Colitis 2017; 11(4): 515.

31.

Ennis

Hinton

Naylor

, et al. A comparison of statistical learning methods on the Gusto database. Stat Med 1998; 17(21): 2501–2508.

32.

van der Ploeg

Austin

Steyerberg

. Modern modelling techniques are data hungry: a simulation study for predicting dichotomous endpoints. BMC Med Res Methodol 2014; 14: 137.

33.

, et al. Development and validation of novel models for the prediction of intravenous corticosteroid resistance in acute severe ulcerative colitis using logistic regression and machine learning. Gastroenterol Rep 2022; 10: goac053.