Development and Validation of an Ileocolonoscopy-Based Nomogram for Predicting Perianal Penetrating Complications in Patients with Crohn’s Disease

Introduction

Crohn’s disease (CD) is a chronic gastrointestinal inflammatory disorder. Its perianal penetrating complications (PPC)—including anal fistula and perianal abscess—are debilitating that manifest as swelling, pain, and discharge of pus, feces or blood. More worryingly, patients with PPC have a threefold risk of anorectal malignancy and higher mortality compared with the general population.^1,2 Surveys indicated the incidence of PPC among CD patients varies from 39.2% to 56.1%.³ However, PPC has an insidious onset, with clinicians usually conducting examinations and adjusting treatment regimens only after overt symptoms emerge; this reveals substantial gaps in PPC monitoring and management, consistent with concerns highlighted in the recently updated European Crohn’s and Colitis Organization (ECCO) guidelines.⁴

Colonic and rectal involvement is a well-established risk factor for PPC in CD, while isolated ileal disease exerts a protective effect; this anatomical association is the core rationale for our study to identify segment-specific ileocolonoscopic features for PPC prediction.⁵ Existing PPC prediction models for CD are mostly based on serum biomarkers⁶ or imaging indicators,⁷ with few models focusing on Ileocolonoscopic features, which are closely linked to PPC pathogenesis.

Ileocolonoscopy serves as the first-line approach for CD diagnosis and The Simple Endoscopic Score for CD (SES-CD) and CD Endoscopic Index of Severity (CDEIS) are critical for evaluation of prognosis. Studies confirmed that reduced scores correlate strongly with better long-term outcomes, including lower risks of perianal lesions.^8,9 Nevertheless, these scoring systems have clinical limitations: scores require segmental calculation with equal weighting assigned to each intestinal segment, even though involved intestinal sites are closely correlated with PPC development.¹⁰ Most importantly, it remains unclear which ileocolonoscopic features—whether included in these systems or not—are predictive risk factors for PPC.

PPC is a key phenotype of transmural penetrating CD, along with upper gastrointestinal penetrating lesions, entero-enteric fistulas, and postoperative anastomotic complications. These entities share chronic transmural inflammation as a common pathological basis, and mucosal lesions seen on ileocolonoscopy serve as direct biomarkers of the inflammatory burden driving penetrating disease. However, few studies have explored the value of ileocolonoscopic mucosal features in predicting PPC.

Thus, we incorporated demographic and ileocolonoscopic characteristics into the study and constructed a straightforward nomogram to predict PPC of CD patients. It also exhibited excellent performance in both internal and external validation. Ileocolonoscopy is only a component of comprehensive PPC risk stratification. Complementary modalities including cross-sectional imaging, nutritional status and systemic inflammatory markers also contribute to risk assessment. Our model aims to complement rather than replace them. Our study highlights the novel value of ileocolonoscopy in early identification and intervention of CD complications.

Materials and Methods

This study is in adherence to the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis).¹¹ The completed checklist is provided as Supplementary Figure - TRIPOD Checklist.

Patient Population

Our study was designed as a two-center retrospective study. Approved by the Institutional Review Board, we enrolled patients diagnosed with CD at the First Affiliated Hospital of Anhui University of Chinese Medicine (FAHAUTCM, Center 1) and Anhui Branch of Shuguang Hospital Affiliated to Shanghai University of Chinese Medicine (ABSHASUTCM, Center 2) between January 1, 2012, and July 31, 2025, with patient informed consent requirements waived. The diagnosis of CD was following the ECCO guidelines.⁴

Inclusion criteria: (1) Complete demographic and ileocolonoscopic data within 7 days of baseline; (2) No ileocolonoscopic treatment (including endoscopic balloon dilation, endoscopic stricturotomy, endoscopic fistula treatment, endoscopic abscess drainage, and endoscopic mucosal or lesion management, etc.) at baseline; (3) No concomitant Montreal p phenotype (perianal fistulas and abscesses); (4) A minimum 3-month follow-up. Exclusion criteria: (1) Age of onset < 14 years; (2) Suboptimal ileocolonoscopic imaging quality; (3) Fewer than 4 intestinal segments available for valid exploration (the segments to be examined include the terminal ileum, ascending colon, transverse colon, descending colon, and rectum, with an insertion depth of at least 10 cm defined as valid exploration of the corresponding intestinal segment)¹² due to various causes (including impassable stenosis, intestinal resection, etc).; (4) Comorbidity with gastrointestinal malignancy, infectious diseases, or other intestinal inflammatory disorders (ulcerative colitis, intestinal tuberculosis, intestinal Behçet’s disease, etc).

Sample size estimation was performed following the approach outlined by Riley et al¹³ The prevalence of PPC among Asian CD patients ranges from 9.3%¹⁴ to 61%.¹⁵ A total of 21 candidate parameters were included with a shrinkage factor of 0.9 and an expected Cox-Snell R² of 0.25. The calculated sample size was 793 cases when the prevalence was set at 9.3%, and 646 cases when the prevalence was set at 61%.

Candidate characteristics including demographic and ileocolonoscopic characteristics were selected based on a synthesis of expertise and literature. The demographic characteristics included gender, age, drinking history, smoking history, disease duration (time from symptom onset to CD diagnosis) and body mass index (BMI). The ileocolonoscopic characteristics documented included the largest ulcer diameter, deep ulcer, edema and hyperemia area, ulcer area, nodular lesions, stenosis and intestinal fistula. Patients enrolled from Center 1 were randomly distributed to a training set and an internal validation set at a 7:3 ratio. Patients’ recruitment flow and grouping strategy are shown in Figure 1.

Figure 1 Flowchart of patient recruitment and grouping.

Baseline and Endpoint

CD diagnosis was defined as baseline, and PPC occurrence as the endpoint. Patients were censored at loss to follow-up or study end (November 30, 2025). Only those with follow-up >3 months were included using their last observed status; the others were excluded for not meeting the minimum follow-up requirement. PPC diagnosis required confirmation by clinical manifestations, digital rectal examination, and imaging.⁴

Ileocolonoscopic Examination

Bowel preparation: Following the Chinese Guidelines for Bowel Preparation,¹⁶ patients were prescribed a low-residue diet for 72 hours pre-examination, with only a clear liquid diet permitted on the day prior. 1L of polyethylene glycol and electrolyte oral solution (PEG-ELS) was administered orally within 1 hour on the eve of the examination. Patients fasted on the examination day and ingested the remaining 3 L of PEG-ELS 4–6 hours pre-procedure, at 250 mL intervals every 15 minutes. The final 1 L of PEG-ELS was mixed with 20 mL of dimethicone before administration.

Personnel allocation and blinding method: Ileocolonoscopic examinations at 2 centers were performed by 3 senior endoscopists from Center 1 (≥ 100 CD case experience each): two conducted and documented examinations, the third resolved discrepancies. Crucially, they were blinded to patients’ disease status and other examination findings, with no involvement in later management.

Examination and documentation: Patients assumed a left lateral decubitus position with flexed knees, followed by digital rectal examination to detect strictures, anal fissures, hemorrhoids, and other anorectal anomalies. They were instructed to breathe orally and relax the anal sphincter for scope insertion. The endoscopist gently inserted the scope into the anus, advanced it to the ileocecal valve, and tried to intubate the terminal ileum as deeply as possible. Systematic imaging and recording of all relevant parameters were completed during the procedure.

Management of adverse reactions: The examination was stopped at once in case of patients’ intolerance or adverse events.

Statistical Analysis

Missing data were assessed for all candidate variables prior to model construction. Variables with missing data >20% were excluded. For missing values of 5–20%, multiple imputation was performed under the missing-at-random assumption. Five imputed datasets were generated using chained equations including clinical and outcome variables. Pooled estimates were obtained via Rubin’s rules, and complete-case analysis was used for sensitivity analysis. For missing data <5%, median imputation was used for non-normally distributed continuous variables and mode imputation for ordinal categorical variables.

Continuous data with normal distribution were presented as mean ± standard deviation and compared via independent samples t-test; non-normally distributed continuous data were expressed as Median (IQR) and compared using Mann–Whitney U-test. Categorical data were summarized as n (%) and compared with chi-square test. All continuous variables were standardized (mean = 0, SD = 1) before analysis. Least Absolute Shrinkage and Selection Operator (LASSO) and Boruta algorithm were then used to address high-dimensional multicollinearity and identify high-predictive features.

The events per variable (EPV) ratio was 19.4 (194 incident PPC events and 5 candidate predictors), well above the recommended minimum threshold of 10 for logistic regression-based clinical prediction models,¹⁷ thus effectively reducing overfitting risk. Further measures to guard against overfitting included 10-fold cross-validation during LASSO feature selection, together with independent internal and external validation to assess model generalizability.

Multivariable logistic regression was prespecified as the primary modeling approach, consistent with our objective to develop a clinically interpretable nomogram for individualized absolute PPC risk prediction at initial CD diagnosis. As the gold-standard method for nomogram development, it provides direct absolute risk estimates for bedside risk stratification, with excellent interpretability for the binary PPC endpoint and low overfitting risk.

Before formal modeling, key assumptions of multivariable logistic regression were thoroughly checked. The linearity assumption in the logit for all continuous predictors was verified using Box–Tidwell tests. No multicollinearity was detected among candidate predictors, as all variance inflation factors (VIF) values were less than 2. Detailed processing and coding rules for variables are provided in Supplementary File - Variable Coding Rules and Handling Methods.

Finally, multivariate logistic regression was employed to adjust for confounding factors and screen stable, significant predictors for model construction. Receiver Operating Characteristic (ROC) curves were constructed to evaluate the model’s discriminative ability, with the Area Under the ROC Curve (AUC) and 95% confidence intervals (CIs) reported. Calibration curves were plotted to assess the consistency between predicted and observed PPC risks, with calibration slopes calculated to quantify goodness of fit; the Hosmer–Lemeshow test was additionally performed for calibration assessment. Decision Curve Analysis (DCA) and nomogram score distribution analysis were used to validate the model’s clinical utility and diagnostic performance. A two-tailed t-test was used, with P < 0.05 considered statistically significant. Statistical analyses were performed using SPSS 24 and R software (http://www.R-project.org).

Result

Baseline Characteristics

431 patients were enrolled in Center 1. No variables had a missing rate exceeding 20%, so no variables were excluded from the analysis. Only two variables had missing data: edema and hyperemia area in the terminal ileum, and ulcer area in the terminal ileum, with an overall missing rate of 2.6% in the Center 1 cohort and 2.4% in the Center 2 cohort (all <5%). Table 1 presents the demographic and ileocolonoscopic characteristics of enrolled patients in the two centers, and Table 2 shows those of patients in the two subgroups of Center 1. Besides, baseline medications and Montreal classification of disease location and behavior were reported in Supplementary File - Supplementary Table. Figure 2 displays representative ileocolonoscopic images of characteristics.

Table 1 Comparison of Demographic and Ileocolonoscopy Characteristics Between Enrolled Patients in Center 1 and Center 2

Table 2 Comparison of Demographic and Ileocolonoscopy Characteristics Between Patients in Training Set and Internal Validation Set

Figure 2 Representative ileocolonoscopic images of typical characteristics. (A) Deep and large ulcer. (B) Mucosal edema and hyperemia. (C) Nodular lesions (cobblestone sign). (D) Stenosis (cannot pass). (E) Intestinal fistula.

Feature Selection

Boruta algorithm and LASSO regression were used for feature screening. Boruta identified 7 noteworthy features from all variables (Figure 3A). LASSO regression with 10-fold cross-validation determined the optimal model with 8 variables (Figure 3B and C). The intersection of them was used to identify features for multivariate logistic regression: the largest ulcer diameter, ulcer area in rectum, ulcer area in descending colon, nodular lesions, stenosis.

Figure 3 Feature screening via Boruta algorithm. (A) Feature screening via Boruta algorithm. Green bars represent confirmed significant variables (n = 7); blue bars represent shadow variables (shadowMin, shadowMax, shadowMean), which are internal negative controls to assess feature importance. (B) Selection of the bias-adjusted parameter (λ) in LASSO regression, with x-axis labeled as Log(λ). (C) Coefficient profile plot based on the log(λ) sequence. Features were selected according to the 1-SE rule (n = 8), with x-axis labeled as Log(λ).

Model Establishment

The 5 features from the intersection were included in multivariate logistic regression to develop the predictive model (Figure 4). The resulting model equation is as follows:

Figure 4 Forest plot of multivariate logistic regression analysis. The horizontal lines are 95% confidence intervals (CIs), and the red squares denote the odds ratios (ORs). The vertical dashed line shows the null effect (OR=1). Variables with 95% CIs that do not cross the null line are considered statistically significant (P<0.05).

Logit(P(PPC)) = −5.579 + 0.408 × (largest ulcer diameter, cm) + 1.065 × (rectal ulcer area, ordinal category) + 0.338 × (descending colon ulcer area, ordinal category) + 0.681 × (nodular lesions, ordinal category) + 0.933 × (stenosis, ordinal category).

Model Performance Evaluation

ROC curves, calibration curves, DCA, and nomogram score distribution were employed to appraise model performance. In the training set, the model achieved an AUC of 0.892 (95% CI: 0.855–0.928), with a sensitivity of 0.720 and specificity of 0.902. The internal validation set showed an AUC of 0.857 (95% CI: 0.794–0.921), sensitivity of 0.795, and specificity of 0.809. The external validation set had an AUC of 0.847 (95% CI: 0.775–0.919), sensitivity of 0.700, and specificity of 0.851 (Figure 5A–C). The slight decrease in AUC is an expected and clinically acceptable finding, suggesting that there may be no obvious overfitting and supporting the model’s good generalizability. Good calibration was observed across all sets, with Hosmer–Lemeshow test P-values of 0.836, 0.240, and 0.498 and calibration slopes of 1.000, 0.843, and 0.845 for the training, internal validation, and external validation sets, respectively, indicating that the model had well goodness of fit (Figure 5D–F). The model yielded positive net benefits when threshold probabilities ranged from 0.01–0.90, 0.01–0.85, and 0.05–0.78 for the three sets (Figure 6A–C). There is a significant difference in nomogram score distribution between PPC and non-PPC groups (P<0.01), confirming good diagnostic performance (Figure 6D–F).

Figure 5 ROC curves and calibration curves. (A) Training set: AUC = 0.892 (95% CI: 0.855–0.928), sensitivity = 0.720, specificity = 0.902. (B) Internal validation set: AUC = 0.857 (95% CI: 0.794–0.921), sensitivity = 0.795, specificity = 0.809. (C) External validation set: AUC = 0.847 (95% CI: 0.775–0.919), sensitivity = 0.700, specificity = 0.851. (D), (E and F) are calibration curves for the three sets. Hosmer–Lemeshow and Spiegelhalter Z-tests, P > 0.05.

Figure 6 DCA and nomogram score distribution boxplots. (A and B) and (C) are DCA of model in 3 sets, respectively. x-axis = threshold probability, y-axis = net benefit. Red line = model-identified CD patients with PPC; All/None lines = extreme strategies of all/none CD patients with PPC. Positive net benefits across thresholds (0.01–0.90, 0.01–0.85, 0.05–0.78) confirm good clinical utility. (D–F) are nomogram score distribution boxplots in 3 sets, respectively. Pink dots = CD patients without PPC; cyan dots = CD patients with PPC. Boxes = 75th percentile (top), median (middle), 25th percentile (bottom). Four asterisks above 2 boxes = significant between-group distribution difference (P < 0.05).

To ease intuitive interpretation, a nomogram was constructed and proved with a clinical case (Figure 7A and B). For clinical utility, an online calculator is freely accessible at https://zsy2001.shinyapps.io/DynNomapp/.

Figure 7 Static-dynamic nomogram. (A) Static nomogram for predicting the risk of PPC in CD patients. (B) Dynamic nomogram for predicting the risk of PPC in CD patients. * denotes P < 0.05, indicating the predictor has a statistically significant association with PPC occurrence; *** denotes P < 0.001, indicating the predictor has an extremely statistically significant association with PPC occurrence.

This study makes an incremental contribution to CD PPC risk prediction. Our work formalizes these routinely collected ileocolonoscopic features into a quantitative, easy-to-use nomogram with validated performance across two independent clinical centers. The model’s discriminative performance (AUC 0.847–0.892) is consistent with that of other well-validated clinical prediction models for IBD-related complications reported in the literature.¹⁸

Discussion

We identified the largest ulcer diameter, ulcer area in rectum, ulcer area in descending colon, nodular lesions and stenosis as predictors of PPC in CD patients and constructed a risk prediction model accordingly. All of predictors were derived from observations of effectively explored intestinal segments, cutting the need for complex calculations and assumption of indetectable segments. Our study shows that features extracted from ileocolonoscopy can effectively predict PPC occurrence with high discrimination ability.

Both SES-CD and CDEIS include ulcers as reliable scoring items. Ma et al¹⁹ developed EASE-CD—a continuous CD activity measurement tool based on ulcer diameter and area ratio with promising performance. Pathogenetically, both ulcers and PPC stem from transmural inflammation. TNF-α, IL-6, IL-1βinduces macrophages, neutrophils, and fibroblasts to secrete MMPs, damaging the mucosal basement membrane and driving ulcer progression to the submucosa, muscularis, and serosa.²⁰ In the perianal region, the cytokine-induced EMT is critical for fistula and abscess formation. EMT-transformed intestinal epithelial cells invade deep tissues, form tubular structures, and connect to other organs or the body surface.²¹ Additionally, dysbiosis may also contribute to fistula formation.²² Jain et al²³ found that Debaryomyces hansenii is enriched in the intestinal tissues of CD, potentially impairing mucosal healing; this dysregulated tissue repair capacity may contribute to the pathogenesis of PPC.²⁴ Our study showed the largest ulcer diameter, ulcer area in rectum and descending colon as independent risk factors for PPC.

Colonic and rectal involvement has been identified as risk factors for PPC,^25,26 while ileal involvement and isolated small intestinal lesions exert a protective effect.²⁷ This association is attributable to anatomical features: first, the colorectum is closer to the perianal region than the proximal intestine, rendering it more susceptible to inflammatory invasion; second, the loose areolar spaces, sphincter complexes, and glandular tissues around the distal rectum and anal canal may accelerate the formation of transmural penetration. Notably, SES-CD and CDEIS assign equal weight to lesions regardless of intestinal segments or types.¹² Our study addressed this limitation and confirmed that ulcer in the rectum and descending colon—not edema and hyperemia, or ulcers in other segments—increase PPC risk, with the risk escalating as the ulcer area ratio expands.

Besides ulcers, stenosis is incorporated as items in scoring systems yet conflicting findings exist on the stenosis-PPC association. The Korean CONNECT study did not find substantive discrepancy in PPC prevalence between participants with and without stenosis,²⁸ while the Japanese iCREST-CD study linked stenosis to a lower incidence of PPC.²⁹ However, U.S. and French studies reported opposite results.^27,30 Despite sharing an inflammatory basis, PPC and stenosis differ in cellular composition: TGF-β-regulated myofibroblasts drive stricturing fibrosis,^31,32 whereas M1 macrophages and Th17/Th1/Th17/1 T cells predominate in fistula linings and adjacent tissues.^21,33 However, EMT may also mediate stenosis formation, as TGF-β triggers intestinal epithelial EMT to upregulate ECM-related protein expression,³⁴ and EMT inhibition preserves intestinal epithelial features.³⁵ Our study confirmed stenosis as a PPC risk factor and further clarified the risk contribution of different stenosis severities.

Nodular lesions are a characteristic endoscopic feature of CD yet absent from existing scoring systems, while the cobblestone sign is one of its preliminary diagnostic criteria.^4,36 Their formation relies on mucosal hyperplasia and ulcer distribution: residual mucosal islands between CD intestinal ulcers protrude under inflammatory stimulation, contrasting with adjacent ulcer defects to form visible nodular lesions; clustered continuous ulcers make these nodules endoscopically identifiable as the cobblestone sign.³⁶ No direct link between nodular lesions and perianal disease has been found, but such lesions show progressive intestinal inflammation—from initial transmural inflammation to mucosal hyperplasia and thickening induced by chronic inflammation, and finally to ulcer margin consolidation via fibrotic repair.^31,37 We tentatively included this characteristic and confirmed that it is a risk factor for PPC. Further validation in larger, diverse populations is warranted, given this study’s retrospective two-center design.

Besides the disease behavior discussed above, B3 (penetrating lesions such as intestinal fistula or intra-abdominal abscesses) and p (perianal disease such as fistula or abscess) are key Montreal classifications in CD and closely associated with PPC development. Patients with baseline p phenotype were excluded at recruitment, eliminating their confounding effect upfront. Intestinal fistula was included as an endoscopic variable but not retained by LASSO and Boruta feature selection. Nevertheless, intestinal fistula and other abdominal penetrating lesions may still confound PPC occurrence, given shared pathophysiological mechanisms including transmural inflammation, EMT, and anatomical proximity.³⁸

Biologics, immunomodulators, 5-ASA agents, and glucocorticoids are core treatments for CD that suppress transmural inflammation and reduce complications including PPC. Infliximab is recommended as first-line therapy for PPC remission.⁴ A recent UK PROFILE trial demonstrated that early infliximab plus immunomodulator significantly reduces fistulas, strictures, and surgeries.³⁹ Thus, medication exposure represents an important potential confounder for PPC risk prediction. In this study, baseline medications were fully reported. Exclusion of medication data from analysis was based on rigorous methodological and clinical considerations. As a long-term retrospective study, medication regimens varied substantially according to disease activity and treatment response. Heterogeneity was further increased by the 2023 inclusion of infliximab in the Chinese national reimbursement drug list, which altered access, adherence, and maintenance therapy. Static baseline data cannot reflect dynamic treatment changes or poor adherence, potentially introducing bias and reducing model stability. Our primary aim was to develop a practical prediction tool using routine, objective, and readily available ileocolonoscopic features for real-time point-of-care risk stratification.

Clinical Translation and Real-World Implementation

This nomogram is designed to complement clinical risk stratification for PPC in CD. Our nomogram offers clinical value in following scenarios: pre-emptive risk stratification for asymptomatic patients, for those with high nomogram scores but no overt PPC, the model supports early treatment escalation before irreversible disease develops; risk refinement for intermediate-risk patients, it can further stratify patients with moderate endoscopic activity into high- and low-risk subgroups to guide personalized management; and integration with multi-modal assessment, it can be combined with conventional risk factors and pelvic MRI to enhance comprehensive risk stratification.

Limitation

First, although multivariable logistic regression was selected for its high clinical interpretability and suitability for nomogram construction, this approach cannot accommodate censored data or variable follow-up durations, which may introduce classification bias. Second, the required sample size was calculated as 646–793 based on a Cox–Snell R² of 0.25, so the model carries a potential risk of overfitting. However, the EPV ratio (19.4; 194 events, 5 predictors) exceeded the recommended threshold of 10, supporting adequate sample size. Multiple strategies, including 10-fold cross-validation in LASSO selection and internal/external validation, were used to minimize overfitting. Third, baseline and follow-up medications were not included in the model; however, we reported baseline medications, evaluated their confounding effects, and justified exclusion of these variables. Fourth, we considered the potential confounding effect of Montreal B3 phenotype. Fifth, this single-country study may have limited generalizability to other populations and settings, requiring further external validation. Other limitations include potential selection bias, incomplete visualization of the proximal small bowel, operator-dependent measurement variability, and restricted applicability in patient’s ineligible for ileocolonoscopy. Future studies will integrate multi-modal data to improve model performance and clinical utility.

Conclusions

This study identified several ileocolonoscopic features associated with PPC in CD and developed a corresponding nomogram. The nomogram showed favorable discrimination, calibration, and clinical utility in two Chinese tertiary hospitals. Consistent with ECCO initiatives, this work facilitates early complication surveillance and personalized management of CD. High-risk patients may benefit from intensified monitoring and early aggressive anti-inflammatory therapy, though prospective trials are needed to validate such preventive strategies. Future studies will evaluate its performance in predicting treatment response, surgical outcomes, and severe disease courses. Further external validation in geographically and ethnically diverse cohorts, as well as integration with cross-sectional imaging and biomarker-based models, is required to confirm its generalizability.