Early Prediction of Primary Non-Response to Infliximab in Crohn’s Disease Using Routinely Available Clinical and Week 2 Laboratory Indicators: A Two-Center Real-World Cohort Study

Abstract

Purpose

Primary non-response (PNR) occurs in a subset of patients with Crohn’s disease (CD) after infliximab (IFX) induction therapy and remains challenging to identify early. This study aimed to develop and perform preliminary external validation of an early PNR risk stratification model using routinely available clinical and Week 2 laboratory indicators.

Patients and Methods

This two-center real-world retrospective cohort study included patients with CD receiving IFX induction therapy from January 2019 to September 2025. Center 1 served as the derivation cohort and Center 2 as the external validation cohort. Model 2C, incorporating complicated disease behavior, Week 2 C-reactive protein (CRP), and Week 2 albumin, was developed using multivariable logistic regression. Model performance was assessed using discrimination, calibration, and overall prediction accuracy. Bootstrap internal validation was performed in the derivation cohort, and the model formula and prespecified fixed cutoff were applied unchanged to the external validation cohort.

Results

Among 201 patients, 63 experienced PNR and 138 did not. The derivation and external validation cohorts included 125 and 76 patients, respectively; complete-case analyses of Model 2C included 120 and 76 patients. Higher Week 2 CRP and lower Week 2 albumin were associated with increased PNR risk. In the derivation cohort, Model 2C showed an apparent AUC of 0.814 and a bootstrap-corrected AUC of 0.797. In the external validation cohort, discrimination remained good (AUC, 0.888), but calibration was suboptimal, with a calibration slope of 0.662 and an observed-to-expected ratio of 0.713, indicating systematic risk overestimation (observed event rate, 34.2%; mean predicted risk, 48.0%). Using the prespecified cutoff of 0.2175, the low-risk group had a PNR rate of 3.4% and a negative predictive value of 96.6%.

Conclusion

Model 2C provides an exploratory approach for early PNR risk stratification during IFX induction therapy, particularly for identifying patients at lower risk of PNR. However, it should not be interpreted as a tool for individual absolute-risk prediction or used as a standalone basis for treatment intensification or switching. Further multicenter prospective validation, recalibration assessment, and clinical-impact studies are required.

Introduction

Crohn’s disease (CD) is characterized by transmural intestinal inflammation and often follows a relapsing course, with substantial clinical heterogeneity across patients. Its pathogenesis is generally considered to arise from the interaction of host susceptibility, microbial imbalance, epithelial barrier dysfunction, and immune dysregulation.^1,2 Infliximab (IFX), an anti-tumor necrosis factor (TNF) monoclonal antibody, remains a widely used biologic option for moderate-to-severe CD, with established roles in remission induction, mucosal healing, and maintenance of disease control.^3,4 However, treatment response to IFX varies substantially among individuals, and a subset of patients shows insufficient improvement during induction despite receiving standard IFX therapy, a condition commonly described as primary non-response (PNR). Reported PNR rates after anti-TNF therapy vary widely, partly because studies differ in outcome definitions, timing of assessment, and the use of objective inflammatory evidence.^5,6 Delayed recognition of PNR may prolong ineffective treatment exposure and delay comprehensive reassessment of disease activity and therapeutic strategy.⁶

Contemporary inflammatory bowel disease management increasingly follows a treat-to-target (T2T) approach, in which symptoms are interpreted together with objective markers of inflammation, including biomarkers, endoscopy, and imaging.⁷ In CD, symptoms do not always correspond to the objective inflammatory burden. Current guidelines highlight the value of biomarkers, such as C-reactive protein (CRP) and fecal calprotectin, in monitoring treatment response and assessing objective inflammation.^8,9 Therefore, early identification of patients at risk of PNR during induction therapy represents an important issue in T2T practice. Week 2 represents an early reassessment point after initial IFX exposure and before the Week 14 outcome assessment window, allowing evaluation of residual inflammatory activity and early treatment response using routinely available laboratory indicators. However, in real-world retrospective studies, induction-period endoscopy, fecal biomarkers, and therapeutic drug monitoring data are often incomplete or selectively available, creating practical challenges for both PNR definition and early risk identification.

In this two-center real-world retrospective cohort, PNR was adjudicated by combining clinical non-response, objective inflammatory evidence, and treatment failure events, with the primary assessment window set at Week 14 ± 2 weeks. We focused on routinely available clinical and laboratory indicators obtainable before treatment and at Week 2, aiming to develop and perform preliminary external validation of a parsimonious, interpretable, and clinically accessible early risk stratification model. The potential value of this model lies mainly in identifying patients at lower risk of PNR and supporting risk-informed clinical reassessment, rather than serving as a standalone threshold for treatment intensification or treatment switching.

Patients and Methods

Study Design and Data Sources

This study was designed as a two-center, retrospective, real-world cohort study. Consecutive patients with Crohn’s disease (CD) who initiated infliximab (IFX) induction therapy between January 2019 and September 2025 at two university-affiliated tertiary hospitals were screened. Eligible patients with assessable PNR outcomes constituted the final study cohort. Patients from the Second Affiliated Hospital of Zhengzhou University constituted the derivation cohort (Center 1), whereas those from the First Affiliated Hospital of Zhengzhou University constituted the external validation cohort (Center 2). The patient screening and cohort allocation process is summarized in Figure 1.

Figure 1.

Patient enrollment flowchart. Patients with Crohn’s disease who initiated infliximab induction therapy between January 2019 and September 2025 were screened. After applying the prespecified eligibility criteria, 201 outcome-assessable patients were included in the main analysis set. Center 1 served as the derivation cohort, and Center 2 served as the external validation cohort.

Abbreviations: CD, Crohn’s disease; IFX, infliximab; PNR, primary non-response.

Study variables were extracted from outpatient and inpatient records, laboratory databases, endoscopic and imaging reports, treatment documentation, and follow-up information. All data were de-identified before analysis. Ethics approval was obtained from the ethics committees of both participating centers (approval numbers: KY2025457 and 2026-KY-0285-002). Because this was a retrospective observational study, the requirement for informed consent was waived by both ethics committees. Prespecified candidate model comparison, final model fitting, cutoff derivation, and bootstrap internal validation were performed in Center 1. Center 2 provided an independent external validation cohort and was not used for predictor selection, model refitting, recalibration, or cutoff re-optimization.

Study Population

Patients were eligible if they met all of the following criteria: (1) age ≥18 years; (2) a confirmed diagnosis of CD based on an integrated assessment of clinical manifestations, laboratory tests, endoscopy, imaging, and histopathology; (3) initiation of IFX induction therapy primarily for active luminal CD; (4) no previous exposure to any biologic agent or targeted small-molecule therapy, with IFX being the first biologic agent used; and (5) availability of sufficient clinical, laboratory, treatment, and follow-up information for baseline assessment, IFX administration evaluation, and PNR outcome adjudication. Active luminal CD was required to be supported by clinical symptoms and objective inflammatory evidence, including elevated inflammatory markers, endoscopically active inflammation or ulceration, or active intestinal wall inflammation on computed tomography enterography, magnetic resonance enterography, or intestinal ultrasound. Patients with concomitant perianal disease could be included, provided that clear evidence of active luminal CD was also present. No lower limit for the baseline Crohn’s Disease Activity Index (CDAI) was prespecified in this study, in order to reflect real-world clinical practice.

Patients were excluded if any of the following conditions applied: (1) IFX treatment initiated primarily for purposes other than the control of active luminal CD, such as isolated perianal disease, postoperative prophylactic therapy, or clinical situations dominated by non-inflammatory lesions; (2) previous treatment with other biologic agents or targeted small-molecule therapies; (3) insufficient evidence supporting the diagnosis of CD or the presence of an alternative diagnosis; (4) active infection, active malignancy, or other serious comorbidities before or at IFX initiation that could affect treatment safety, interpretation of inflammatory markers, or attribution of treatment response; (5) pregnancy or lactation; (6) discontinuation of IFX for non-efficacy-related reasons, including infection, infusion reaction, intolerance, adherence issues, financial reasons, administrative reasons, or loss to follow-up, when the PNR outcome could not be reliably adjudicated; or (7) substantial missingness in key baseline characteristics, IFX administration information, or outcome adjudication data that prevented reliable study inclusion or PNR adjudication. If sufficient evidence fulfilling the prespecified PNR criteria was already available before discontinuation for non-efficacy-related reasons, the patient remained eligible for analysis. Cases with unclear reasons for treatment adjustment or discontinuation were independently reviewed and adjudicated by two investigators, with disagreements resolved by a third investigator.

Diagnosis, IFX Treatment, and Assessment Windows

Gastroenterologists confirmed the diagnosis of CD by integrating clinical, laboratory, endoscopic, imaging, and histopathological findings, with reference to relevant ECCO guidelines for diagnosis and monitoring.¹⁰ Disease phenotype was categorized according to the Montreal classification, including age at diagnosis, disease location, disease behavior, and perianal disease. Complicated disease behavior was defined as stricturing (B2) or penetrating (B3) disease. Clinical activity was assessed primarily using the CDAI. When endoscopic data were available, mucosal inflammatory activity was evaluated using the SES-CD. CTE, MRE, or intestinal ultrasound was used to provide complementary evidence of objective inflammatory activity, when available.¹¹

IFX induction consisted of intravenous infusion at 5 mg/kg at Weeks 0, 2, and 6, consistent with standard practice.¹² The first infusion date was defined as Day 0, corresponding to Week 0. Baseline clinical and laboratory data were preferentially extracted from records obtained on Day 0 before drug administration. If data on that day were unavailable, the record closest to Day 0 within 2 weeks before infusion was used. If no suitable record within this interval was available, the most recent record within 4 weeks before infusion was considered if it corresponded to the same clinical assessment phase as IFX initiation. Endoscopic and imaging data were obtained from the closest auditable examination within 8 weeks before the first IFX infusion. All baseline variables were derived from data collected before the first IFX infusion.

The Week 2 window was defined as Day 14 ± 3 days. Clinical assessments and laboratory indicators were preferentially extracted before the second IFX infusion or on the day of infusion before drug administration. If the sequence relative to infusion could not be clearly determined, the record closest to Day 14 within this window and corresponding to the same clinical assessment was used. The primary outcome assessment window was Week 14 ± 2 weeks, corresponding to Days 84–112. The observation window for treatment failure events was Days 0–112. For baseline, Week 2, and Week 14 assessments, if multiple records were available within the same time window, the record closest to the target date and corresponding to the same clinical assessment was preferentially selected. Measurements within the same assessment window were not averaged.

Variable Definitions and Data Processing

Demographic characteristics, disease phenotype, laboratory indicators, and treatment-related information were collected according to prespecified time points and assessment windows. Demographic and disease-related variables included sex, age, body mass index (BMI), smoking history, alcohol use, disease duration, Montreal classification, perianal disease, extraintestinal manifestations, CD-related complications, and previous surgical history. Laboratory indicators included inflammatory, nutritional, hematological, and coagulation-related parameters at Week 0 and Week 2. Among these, CRP and albumin were prespecified candidate variables considered for the main model because of their clinical relevance and routine availability. Laboratory units were standardized across centers, and center-specific reference ranges were recorded. For CRP values reported below the lower limit of detection, values were assigned as one-half of the reported lower limit for analysis.

Changes from baseline to Week 2 were defined as the Week 2 value minus the baseline value and were calculated only when both time points were available. The percentage reduction in CRP was defined as (baseline CRP − Week 2 CRP) / baseline CRP × 100% and was calculated only in patients with baseline CRP above the center-specific upper reference limit, baseline CRP >0, and available Week 2 CRP. To reduce collinearity, the baseline value, Week 2 value, and change value of the same indicator were generally not entered simultaneously into the same main model. Treatment-related information included IFX dose, actual administration dates, treatment adjustment or switching and the corresponding reasons, as well as concomitant immunomodulators, corticosteroids, enteral nutrition, and anti-infective therapy during induction. Therapeutic drug monitoring data were collected when available but were not included in the main model because testing was selectively performed in routine practice and available measurements occurred at heterogeneous time points. Week 14 assessment variables were used only for outcome adjudication and were not considered early predictors.

Outcome Definition and Adjudication Quality Control

The primary outcome of this study was PNR after IFX induction therapy. The main analysis used a binary outcome of PNR versus non-PNR. The primary time point for PNR adjudication was Week 14 ± 2 weeks. For treatment failure events occurring within Days 0–112 that could be clearly attributed to insufficient IFX efficacy or progression of active CD, the event date was used as the outcome anchor time point.

Patients were classified as having PNR when either of the following two alternative adjudication pathways was fulfilled. No numerical weighting or hierarchical priority was assigned between these pathways. First, clinical non-response at Week 14 ± 2 weeks accompanied by persistent objective inflammatory activity. Clinical non-response was defined as a decrease in CDAI of <70 points from baseline at Week 14 ± 2 weeks and a CDAI >150 at Week 14 ± 2 weeks. Persistent objective inflammatory activity was defined as follows: among patients with baseline CRP above the center-specific upper reference limit, CRP remained above the reference limit at Week 14 ± 2 weeks and decreased by <50% from baseline; or, among patients with both pretreatment and Week 14 ± 2 weeks SES-CD data available, SES-CD decreased by <50% from baseline and remained >2 at Week 14 ± 2 weeks; or contemporaneous CTE, MRE, or intestinal ultrasound indicated persistent active inflammation or no clear improvement from baseline. These criteria represented availability-dependent objective inflammatory evidence and were not mutually exclusive. Second, a treatment failure event occurred within Days 0–112 and was confirmed by medical record review to be attributable to insufficient IFX efficacy or progression of active CD. Such events included IFX dose intensification clearly documented as a response to inadequate efficacy, discontinuation of IFX and switching to another biologic agent or targeted small-molecule therapy, initiation or substantial dose escalation of systemic corticosteroids, hospitalization for intensified treatment of active CD, and CD-related surgery, abscess drainage, or other interventional procedures. Adjustments unrelated to efficacy, or those with uncertain attribution, were not counted as treatment failure events.

Non-PNR was defined as not fulfilling the PNR criteria and having no treatment failure event attributable to insufficient IFX efficacy or progression of active CD within Days 0–112. Clinical response was defined as a decrease in CDAI of ≥70 points from baseline, and clinical remission was defined as CDAI ≤150. Cases with uncertain or non-evaluable outcomes were not included in the primary analysis set or model development.

CDAI was retrospectively reconstructed according to its standard components using previous medical records, including outpatient records, inpatient progress notes, inflammatory bowel disease follow-up records, nursing assessment forms, and contemporaneous laboratory tests. Two investigators independently extracted data and calculated CDAI. Disagreements regarding key components, final scores, PNR outcome adjudication, or attribution of treatment failure events were resolved by a third investigator. The composition of the primary analysis set, derivation cohort, external validation cohort, and complete-case population for Model 2C is shown in Figure 1 and Supplementary Table S2. The outcome variable was not imputed. The main model and candidate models were developed using complete-case analyses according to the variables included in each model.

Model Development and Validation

Before comparing candidate models, predictors were defined according to clinical availability, supporting evidence, timing of measurement, completeness, and biological plausibility. Given the limited number of PNR events in the derivation cohort, model development followed a parsimonious strategy, prioritizing variables that were available at Week 0 and Week 2, clinically interpretable, and had low proportions of missing data. Predictor selection for the main model was not based solely on univariable P values, and automated data-driven selection methods, including the least absolute shrinkage and selection operator (LASSO) and stepwise regression, were not used.

Four prespecified candidate frameworks were evaluated: the Baseline model, Dynamic change model, Week 2 lab-only model, and Model 2C. Within this prespecified candidate framework, the final main model, Model 2C, was selected by integrating clinical interpretability, variable accessibility, missingness, model parsimony, and the overall characteristics of the prespecified candidate models. Model 2C included complicated disease behavior, Week 2 CRP, and Week 2 albumin. Model 2C was developed using multivariable logistic regression. The regression results of Model 2C are shown in Table 1, and the full linear predictor formula and variable coding are provided in Supplementary Table S7A. The corresponding total-point–risk mapping is shown in Supplementary Table S7B.

Table 1.

Multivariable Logistic Regression Results for Model 2C in the Derivation Cohort

Predictor	β coefficient	SE	OR (95% CI)	P value
Complicated disease behavior	0.021	0.498	1.02 (0.39–2.71)	0.966
Week 2 C-reactive protein, per 1 mg/L increase	0.149	0.039	1.16 (1.08–1.25)	<0.001
Week 2 albumin, per 1 g/L increase	−0.119	0.046	0.89 (0.81–0.97)	0.010

Notes: Model 2C comprised complicated disease behavior, Week 2 C-reactive protein, and Week 2 albumin. Odds ratios for C-reactive protein and albumin correspond to each 1 mg/L and 1 g/L increase, respectively. Complicated disease behavior indicates stricturing or penetrating behavior based on the Montreal classification.

Abbreviations: β, beta coefficient; CI, confidence interval; CRP, C-reactive protein; OR, odds ratio; SE, standard error.

Model discrimination was evaluated using the area under the receiver operating characteristic curve (AUC) and corresponding 95% confidence intervals, while overall prediction error was summarized by the Brier score. Calibration was assessed with calibration plots and summarized using the calibration intercept and slope. In the external validation cohort, we additionally calculated the observed event rate, mean predicted risk, and observed-to-expected ratio (O/E ratio). Decision curve analysis (DCA) was performed across threshold probabilities from 0.05 to 0.80 to examine potential net benefit, with treat-all and treat-none strategies used as comparators.

In the derivation cohort, internal validation was performed using 1000 bootstrap resamples to estimate optimism-corrected model performance. In the external validation cohort, the original Model 2C formula and regression coefficients derived from the derivation cohort were directly applied to calculate the predicted probability for each patient and to evaluate model performance. No model refitting, recalibration, predictor re-selection, or cutoff re-optimization was performed. A nomogram and total points–risk mapping were constructed based on the original logistic model for visualization only, and patient-level auditing was performed to confirm consistency with the original model predictions. The fixed cutoff of 0.2175 was derived from the Youden index of the ROC analysis in the derivation cohort and was applied unchanged to the external validation cohort. This cutoff was used only for risk stratification and model interpretation, and not as a direct basis for treatment intensification or treatment switching.

Statistical Analysis

Continuous data were described using means with standard deviations or medians with interquartile ranges (IQR), depending on distributional characteristics. Categorical data were reported as frequencies and percentages. Between-group comparisons were selected according to data type and distribution, including the independent-samples t-test, Welch’s t-test, Mann–Whitney U-test, chi-square test, and Fisher’s exact test where appropriate. All statistical tests were two-sided, and P<0.05 was considered statistically significant. Descriptive between-group comparisons were used only to present differences in characteristics between the PNR and non-PNR groups and were not used as the sole basis for predictor inclusion in the model. All statistical analyses were performed using R software, version 4.3.0.

Calibration plots were constructed using deciles of predicted risk within each cohort. Within each decile, the mean predicted risk was compared with the observed primary non-response event rate, and exact Clopper–Pearson binomial 95% confidence intervals were calculated for the observed event rate. Point size was proportional to the number of patients in each decile.

Reporting Guidelines

This study was reported in accordance with the Reporting of Studies Conducted Using Observational Routinely Collected Health Data (RECORD) statement. A completed RECORD checklist was provided as a supplementary reporting-guideline file. Because this study involved the development and external validation of a prediction model, a completed TRIPOD checklist was also submitted as supplementary reporting-guideline material. The patient enrollment flowchart is presented in Figure 1.

Results

Study Population and Cohort Composition

A total of 241 patients with CD who received IFX induction therapy at the two centers between January 2019 and September 2025 were screened. According to the prespecified inclusion and exclusion criteria, 40 patients were excluded, and 201 outcome-assessable patients were finally included, comprising 63 patients in the PNR group and 138 in the non-PNR group. Center 1 served as the derivation cohort and included 125 patients, including 37 in the PNR group and 88 in the non-PNR group. Center 2 served as the external validation cohort and included 76 patients, including 26 in the PNR group and 50 in the non-PNR group. The patient screening and cohort allocation process is shown in Figure 1.

In the complete-case analysis for Model 2C, 120 patients were included in the derivation cohort, comprising 36 in the PNR group and 84 in the non-PNR group; 76 patients were included in the external validation cohort, comprising 26 in the PNR group and 50 in the non-PNR group. Comparisons of baseline and Week 2 characteristics between the derivation and external validation cohorts are shown in Supplementary Table S1. The complete-case assessment for Model 2C is provided in Supplementary Table S2.

Clinical Characteristics of the PNR and Non-PNR Groups

Patients in the PNR and non-PNR groups differed in several baseline characteristics. Compared with the non-PNR group, the PNR group had a longer disease duration and a different distribution of Montreal disease location. The proportion of complicated disease behavior was higher in the PNR group, although the difference did not reach statistical significance. Regarding baseline laboratory indicators, the PNR group had lower hemoglobin, albumin, and prealbumin levels, while baseline CRP showed an increasing trend. Baseline SES-CD also differed between the two groups; however, the value was lower in the PNR group, and endoscopic data were partly missing.

Week 2 indicators showed clearer between-group differences. Week 2 CRP was significantly higher in the PNR group than in the non-PNR group [14.95 (3.52–32.26) mg/L vs 1.51 (0.54–5.44) mg/L, P<0.001], whereas Week 2 albumin was significantly lower [36.75 (33.80–41.20) g/L vs 41.55 (38.45–44.20) g/L, P<0.001]. These findings were consistent with a higher residual inflammatory burden and less favorable nutrition-related laboratory profiles in the PNR group during early IFX induction therapy. The main clinical characteristic comparison is shown in Table 2. A complete comparison of all available variables, together with the available sample size for each variable, is provided in Supplementary Table S9.

Table 2.

Key Baseline and Week 2 Characteristics According to Primary Non-Response Status

Characteristic	Overall (n = 201)	non-PNR (n = 138)	PNR (n = 63)	P value
Demographic characteristics
Age, years	31.00 (22.00–45.00)	30.50 (21.25–43.00)	35.00 (25.50–46.00)	0.179
Male sex	150 (74.6%)	102 (73.9%)	48 (76.2%)	0.865
Body mass index, kg/m2	19.38 (17.39–22.08)	19.72 (17.58–22.20)	18.82 (16.98–21.30)	0.166
Disease history and phenotype
Disease duration, months	5.00 (1.00–24.00)	4.00 (1.00–14.75)	8.00 (1.50–36.00)	0.045
Montreal disease location				0.015
L1	44 (21.9%)	25 (18.1%)	19 (30.2%)
L2	60 (29.9%)	37 (26.8%)	23 (36.5%)
L3	97 (48.3%)	76 (55.1%)	21 (33.3%)
Upper gastrointestinal involvement (L4 modifier)	27 (13.4%)	18 (13.0%)	9 (14.3%)	0.987
Montreal disease behavior				0.224
B1	113 (56.2%)	83 (60.1%)	30 (47.6%)
B2	69 (34.3%)	44 (31.9%)	25 (39.7%)
B3	19 (9.5%)	11 (8.0%)	8 (12.7%)
Perianal disease	48 (23.9%)	38 (27.5%)	10 (15.9%)	0.105
Complicated disease behavior	88 (43.8%)	55 (39.9%)	33 (52.4%)	0.132
Baseline CDAI	220.76 ± 93.15	214.49 ± 88.79	234.50 ± 101.43	0.181
Baseline SES-CD	11.00 (7.00–18.00)	12.00 (7.00–19.00)	10.00 (6.25–13.75)	0.040
Baseline laboratory indicators
Baseline C-reactive protein, mg/L	16.41 (4.96–43.11)	15.09 (4.56–38.44)	22.61 (7.35–51.73)	0.059
Baseline C-reactive protein above ULN	118 (59.6%)	80 (58.4%)	38 (62.3%)	0.719
Baseline albumin, g/L	39.10 (35.10–43.30)	39.90 (35.40–44.20)	37.90 (31.95–40.98)	0.005
Baseline prealbumin, mg/L	196.61 ± 77.61	205.82 ± 74.41	175.69 ± 81.40	0.031
Week 2 laboratory indicators
Week 2 C-reactive protein, mg/L	3.08 (0.77–11.42)	1.51 (0.54–5.44)	14.95 (3.52–32.26)	<0.001
Week 2 albumin, g/L	40.65 (36.50–43.80)	41.55 (38.45–44.20)	36.75 (33.80–41.20)	<0.001
Treatment-related variables
Concomitant steroid use	13 (6.5%)	11 (8.0%)	2 (3.2%)	0.353
Concomitant immunomodulator use	16 (8.0%)	10 (7.2%)	6 (9.5%)	0.785

Notes: Values are shown as mean ± standard deviation, median (interquartile range), or n (%), according to variable type and distribution. Group comparisons were performed using Student’s t-test, Welch’s t-test, Mann–Whitney U-test, chi-square test, or Fisher’s exact test, as appropriate. Variable-level available sample sizes are listed in Supplementary Table S9.

Abbreviations: BMI, body mass index; CD, Crohn’s disease; CDAI, Crohn’s Disease Activity Index; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; IFX, infliximab; IQR, interquartile range; PNR, primary non-response; SES-CD, Simple Endoscopic Score for Crohn’s Disease; ULN, upper limit of normal.

Comparison of Prespecified Candidate Models

In the derivation cohort, the apparent performance of the prespecified candidate models showed that the baseline model, which included only baseline information, had limited discriminative ability, with an AUC of 0.702 (95% CI 0.598–0.807) and a Brier score of 0.179. The dynamic change model, which incorporated dynamic change indicators, had an AUC of 0.703 (95% CI 0.595–0.811) and a Brier score of 0.178, showing similar apparent performance compared with the baseline model.

By contrast, models including Week 2 laboratory indicators showed higher apparent discrimination. Among single-indicator models, the Week 2 CRP only model had an AUC of 0.764 (95% CI 0.666–0.863), and the Week 2 albumin only model had an AUC of 0.699 (95% CI 0.588–0.811). The Week 2 lab-only model, which combined these two variables, had an AUC of 0.814 (95% CI 0.725–0.903) and a Brier score of 0.142. The final Model 2C, which included complicated disease behavior, Week 2 CRP, and Week 2 albumin, also had an AUC of 0.814 (95% CI 0.725–0.903) and a Brier score of 0.142. These results suggest that the main information contributing to apparent model discrimination came from Week 2 CRP and Week 2 albumin, rather than from baseline information alone or dynamic change indicators.

The addition of complicated disease behavior did not change the apparent AUC or Brier score to three decimal places. However, as a prespecified clinical phenotype variable, it was retained in the main model framework to preserve clinical interpretability. Within the prespecified candidate framework, Model 2C was selected as the final main model for subsequent internal and external validation. Detailed apparent performance of the candidate models and single-indicator reference models is shown in Supplementary Table S3, and the AUC comparison of the main candidate models is shown in Supplementary Figure S2.

Regression Results of Model 2C

Among the 63 patients classified as having PNR, 60 fulfilled the Week 14 clinical non-response with persistent objective inflammatory activity pathway, 3 fulfilled the treatment failure pathway, and no patient fulfilled both pathways (Supplementary Table S10). Model 2C included complicated disease behavior, Week 2 CRP, and Week 2 albumin. In the complete-case derivation cohort, multivariable logistic regression showed that each 1 mg/L increase in Week 2 CRP was associated with an increased risk of PNR, with an OR of 1.16 (95% CI 1.08–1.25, P<0.001). Each 1 g/L increase in Week 2 albumin was associated with a lower risk of PNR, with an OR of 0.89 (95% CI 0.81–0.97, P=0.010). The OR for complicated disease behavior was 1.02 (95% CI 0.39–2.71, P=0.966), indicating no statistically significant association in this model.

The multivariable logistic regression results of Model 2C are shown in Table 1, and the ORs and 95% CIs for the predictors are visualized in Figure 2. The complete linear predictor formula is provided in Supplementary Table S7A.

Figure 2.

Forest plot of Model 2C predictors. Predictor effects in Model 2C are shown as odds ratios with 95% confidence intervals in the derivation cohort. Week 2 C-reactive protein and Week 2 albumin were modeled continuously, per 1 mg/L and 1 g/L increase, respectively. The dashed vertical reference line marks an odds ratio of 1.

Abbreviations: CI, confidence interval; CRP, C-reactive protein; OR, odds ratio.

Model Performance, Internal Validation, and External Validation

Model 2C showed good apparent discrimination in the derivation cohort, with an AUC of 0.814 (95% CI 0.725–0.903) and a Brier score of 0.142. Bootstrap internal validation was based on 1000 successful resamples. The bootstrap-corrected AUC was 0.797, the corrected Brier score was 0.154, and the corrected calibration slope was 0.921, suggesting some optimism in the apparent performance estimates but acceptable overall stability.

Five patients in the derivation cohort had missing Week 2 CRP and albumin values and were excluded from the primary complete-case Model 2C analysis. In the multiple imputation sensitivity analysis including all 125 derivation-cohort patients, the pooled estimates remained consistent with the complete-case analysis. The pooled ORs were 1.05 (95% CI 0.39–2.80) for complicated disease behavior, 1.16 (95% CI 1.07–1.25) for Week 2 CRP, and 0.88 (95% CI 0.81–0.97) for Week 2 albumin. Mean apparent performance across the 50 imputed datasets was comparable with the complete-case analysis (AUC, 0.816 vs 0.814; Brier score, 0.141 vs 0.142) (Supplementary Table S11).

In the Center 2 external validation cohort, Model 2C had an AUC of 0.888 (95% CI 0.806–0.969) and a Brier score of 0.144, indicating that the model retained good discrimination in an independent center. However, the calibration intercept was −1.140 and the calibration slope was 0.662 in the external validation cohort. The observed event rate was lower than the mean predicted risk (34.2% vs 48.0%), with an O/E ratio of 0.713, indicating overall risk overestimation in the external validation cohort. DCA showed that Model 2C had potential net benefit across the prespecified threshold probability range of 0.05–0.80. Model performance results are shown in Table 3 and Figure 3, and detailed results of bootstrap internal validation and external validation are provided in Supplementary Tables S4 and S5. The distribution of predicted risks generated by Model 2C is shown in Supplementary Figure S1.

Table 3.

Predictive Performance of Model 2C in the Derivation Cohort, Bootstrap Internal Validation, and External Validation Cohort

Validation step or cohort	n	PNR/non-PNR	AUC (95% CI)	Brier Score	Calibration Intercept	Calibration Slope	Observed Event Rate	Mean Predicted Risk	O/E Ratio
Apparent performance in the derivation cohort	120	36/84	0.814 (0.725–0.903)	0.142			30.0%	30.0%
Bootstrap-corrected performance in the derivation cohort	120	36/84	0.797	0.154		0.921
External validation cohort	76	26/50	0.888 (0.806–0.969)	0.144	−1.140	0.662	34.2%	48.0%	0.713

Notes: Model 2C comprised complicated disease behavior, Week 2 C-reactive protein, and Week 2 albumin. Apparent performance was calculated in complete cases from the derivation cohort. Bootstrap estimates were based on 1000 successful resamples; confidence intervals were not calculated for optimism-corrected indices. External validation used the locked derivation-cohort model without refitting, recalibration, predictor re-selection, or cutoff re-optimization. Empty cells denote unavailable or non-applicable metrics. In the external validation cohort, lower observed event rate than mean predicted risk and an O/E ratio <1 indicated overall risk overestimation. Observed event rate and mean predicted risk are percentages.

Abbreviations: AUC, area under the receiver operating characteristic curve; CI, confidence interval; CRP, C-reactive protein; O/E, observed-to-expected ratio; PNR, primary non-response.

Figure 3.

Predictive performance of Model 2C in the derivation and external validation cohorts. (A) Receiver operating characteristic curves showing model discrimination in the derivation cohort (n=120) and external validation cohort (n=76); the corresponding AUC values are shown within the panel. (B) Calibration plot based on deciles of predicted risk within each cohort, comparing mean predicted risk with the observed primary non-response event rate. Each point represents one risk decile, and point size is proportional to the number of patients in that decile. Each derivation-cohort decile contained 12 patients, whereas the external validation deciles contained 7 or 8 patients. Vertical error bars indicate exact Clopper–Pearson binomial 95% confidence intervals for the observed event rate. (B) the dashed diagonal line indicates ideal calibration. The external validation calibration slope of 0.662 and observed-to-expected ratio of 0.713 are shown within the panel. (C) Decision curve analysis in the derivation cohort. (D) Decision curve analysis in the external validation cohort. Treat-all and treat-none strategies are included as reference strategies in (C and D). (A) The dashed diagonal line indicates no-discrimination performance.

Abbreviations: AUC, area under the receiver operating characteristic curve; CI, confidence interval; DCA, decision curve analysis; O/E, observed-to-expected ratio; PNR, primary non-response.

Fixed-Cutoff Risk Stratification

Using the fixed cutoff of 0.2175 determined in the derivation cohort, patients were stratified into low-risk and high-risk groups. This cutoff was applied unchanged to the external validation cohort and was not re-optimized in that cohort.

In the derivation cohort, the PNR rate was 11.6% (8/69) in the low-risk group and 54.9% (28/51) in the high-risk group. In the external validation cohort, the PNR rate was 3.4% (1/29) in the low-risk group and 53.2% (25/47) in the high-risk group. Under the fixed cutoff, sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were 77.8%, 72.6%, 54.9%, 88.4%, and 74.2%, respectively, in the derivation cohort, and 96.2%, 56.0%, 53.2%, 96.6%, and 69.7%, respectively, in the external validation cohort.

The fixed cutoff separated patients into groups with different observed PNR rates in both cohorts. The low-risk group had a low PNR rate and a high negative predictive value in the external validation cohort, whereas the observed PNR rate in the high-risk group exceeded 50% in both cohorts. The fixed-cutoff risk stratification results are shown in Table 4 and Figure 4, and the confusion matrix is provided in Supplementary Table S6. Observed PNR rates according to the fixed-cutoff risk groups are presented in Table 4A, and classification performance at the same cutoff is summarized in Table 4B.

Table 4.

Fixed-Cutoff Risk Stratification and Classification Performance of Model 2C

A Observed Primary Non-response Rates According to Fixed-cutoff Risk Groups
Cohort	Risk group	n	Observed PNR, n (%)		Observed Non-PNR, n (%)
Derivation cohort	Low-risk group	69	8 (11.6%)		61 (88.4%)
Derivation cohort	High-risk group	51	28 (54.9%)		23 (45.1%)
External validation cohort	Low-risk group	29	1 (3.4%)		28 (96.6%)
External validation cohort	High-risk group	47	25 (53.2%)		22 (46.8%)
B Classification Performance at the Fixed Cutoff
Cohort	Fixed Cutoff	Sensitivity	Specificity	PPV	NPV	Accuracy
Derivation cohort	0.2175	77.8%	72.6%	54.9%	88.4%	74.2%
External validation cohort	0.2175	96.2%	56.0%	53.2%	96.6%	69.7%

Notes: The fixed cutoff of 0.2175 was derived in the derivation cohort using the Youden index and was applied unchanged to the external validation cohort. Predicted risks below this value were assigned to the low-risk group, whereas values at or above this cutoff were assigned to the high-risk group. For classification metrics, high-risk status was considered the positive prediction and primary non-response was considered the positive outcome. Positive and negative predictive values depend on the observed event rate of primary non-response in each analysis cohort—30.0% in the complete-case derivation cohort and 34.2% in the external validation cohort—and should therefore be interpreted in the context of cohort-specific event rates. Direct cross-cohort comparisons of these predictive values should be made cautiously. This cutoff was used for exploratory risk stratification only and should not be interpreted as a standalone threshold for treatment intensification or treatment switching.

Abbreviations: NPV, negative predictive value; PPV, positive predictive value; PNR, primary non-response.

Figure 4.

Fixed-cutoff risk stratification and classification performance. (A) Observed primary non-response rates in low- and high-risk groups based on the fixed cutoff of 0.2175. The cutoff was derived in the derivation cohort and used unchanged in the external validation cohort. (B) Classification metrics at the fixed cutoff in both cohorts. Cutoff-based grouping was intended for risk stratification and should not be interpreted as a standalone threshold for treatment intensification or treatment switching.

Abbreviations: NPV, negative predictive value; PNR, primary non-response; PPV, positive predictive value.

Supplementary Model Visualization and Nomogram Consistency Audit

The distribution of predicted risk showed that the PNR group generally had higher predicted risks than the non-PNR group, although overlap between the two groups was observed. This supports the use of Model 2C as a risk stratification tool rather than a deterministic classification tool (Supplementary Figure S1). To enhance interpretability, a nomogram and total points–risk mapping were constructed for Model 2C. Patient-level checks showed that the nomogram point system was highly consistent with the predictions from the original logistic model, with maximum, mean, and median absolute prediction errors all <1×10⁻⁸, and no patient exceeded the displayed range of the nomogram. These findings indicate that the nomogram served only as a visual interpretation tool and did not alter the original model formula or predicted results (Supplementary Figure S3 and Supplementary Tables S7A, S7B, and S8).

Discussion

Main Findings

Based on a two-center real-world cohort of patients with CD, this study developed and preliminarily validated a parsimonious early risk stratification model for PNR during IFX therapy. The main findings were as follows. Differences between the PNR and non-PNR groups were mainly reflected in residual inflammatory burden and nutrition–inflammation status, as indicated by Week 2 CRP and albumin. Model 2C consisted of complicated disease behavior, Week 2 CRP, and Week 2 albumin, with its predictive information mainly derived from the Week 2 indicators. The model showed good discrimination in both the derivation cohort and the external validation cohort, although external validation indicated a degree of risk overestimation. The fixed cutoff of 0.2175 was particularly useful for identifying patients at low risk of PNR.

It should be emphasized that Model 2C is not intended to replace the existing framework for evaluating IFX treatment response, nor should it be interpreted as a tool for directly determining treatment intensification or treatment switching. Its more appropriate role is to support PNR risk stratification early during IFX induction therapy using routinely available indicators, thereby providing a reference for subsequent reassessment and comprehensive clinical judgment, rather than serving as a deterministic diagnostic tool for PNR or a standalone treatment decision tool.

Relationship with Previous Studies

In previous studies on predictors of response to anti-TNF therapy, the definitions of PNR, assessment windows, and objective inflammatory anchors have not been fully consistent. Some studies have relied on clinical symptoms or changes in disease activity indices, some have emphasized endoscopic improvement, and others have incorporated drug concentrations, anti-drug antibodies, or treatment intensification events.^6,13–16 The present study used a PNR adjudication framework integrating clinical non-response, objective inflammatory evidence, and treatment failure events related to insufficient efficacy. The aim was not to propose a new standard definition of PNR, but to balance clinical relevance and feasibility of outcome adjudication in a real-world setting where data completeness is often limited.

Previous studies have attempted to develop prediction models for IFX-related PNR using routine clinical data. Ye et al developed a PNR nomogram based on a multicenter cohort and found that Week 2 CRP was an important predictor. However, that study defined PNR as a <50% decrease in CDEIS at Week 14 compared with baseline, requiring complete pretreatment and post-induction endoscopic data; therefore, its findings require further validation in real-world patients with incomplete endoscopic reassessment or outcomes requiring integrated adjudication.¹³ Therapeutic drug monitoring, molecular biomarkers, omics-based features, and multimodal models have also provided clues for predicting response to anti-TNF therapy. However, these approaches often rely on additional testing or complex modeling procedures, which may limit their accessibility and generalizability in routine real-world settings.^5,17–19 The purpose of this study was not to replace complex biomarker-based or multimodal models, but to provide a simple and interpretable early risk stratification model based on routinely available indicators, particularly Week 2 CRP and albumin.

Significance of Week 2 Residual Inflammatory Burden and Nutrition–Inflammation Status

The findings of this study highlight the value of Week 2 indicators for early PNR risk stratification during IFX induction therapy. Compared with baseline CRP or baseline albumin, Week 2 CRP and albumin may more closely reflect early inflammatory control and systemic status after treatment initiation. Persistently elevated Week 2 CRP may indicate insufficient inflammatory control or an inadequate early treatment response. Lower Week 2 albumin may reflect ongoing inflammatory consumption, poor nutritional status, intestinal protein loss, or an unfavorable pharmacokinetic background.

Albumin should not be interpreted merely as a nutritional marker in this study. Previous pharmacokinetic studies of IFX have shown that serum albumin, body weight, and anti-IFX antibodies are associated with IFX clearance, and lower albumin may be related to higher drug clearance and lower drug exposure.^20–22 Therefore, lower Week 2 albumin may indicate PNR risk through both nutrition–inflammation status and potential drug exposure. However, because IFX trough concentrations and anti-drug antibodies were not systematically included in this study, we cannot directly demonstrate that low albumin leads to PNR through altered IFX exposure. This potential mechanism requires further investigation.

The candidate model comparison further supports this interpretation. Models including Week 2 CRP and albumin performed better than models relying only on baseline information or dynamic change indicators, suggesting that residual inflammatory burden and nutrition–inflammation status at Week 2 represent core information for early PNR risk stratification. Both CRP and albumin are routinely measured clinical indicators with low cost and high availability, which also enhances the interpretability of the model in real-world settings.

Interpretation of Complicated Disease Behavior

Complicated disease behavior was included in Model 2C as a prespecified clinical phenotype variable, but it did not show a clear independent predictive contribution in the final model. Stricturing or penetrating disease usually represents a more complex disease background and may be clinically related to complication risk, treatment pathways, and subsequent management strategies. However, after Week 2 CRP and Week 2 albumin were included, the incremental predictive information provided by complicated disease behavior in this model was limited. This finding should not be interpreted as indicating that complicated disease behavior is unimportant for the long-term prognosis or management of CD.

Therefore, the value of complicated disease behavior in Model 2C mainly lies in preserving a prespecified clinical phenotype component and maintaining clinical interpretability of the model framework. It should not be overinterpreted as a strong predictor. Interpretation of the model should focus primarily on Week 2 CRP and Week 2 albumin rather than on complicated disease behavior alone.

External Validation, Calibration, and Clinical Implications of the Fixed Cutoff

Derivation-cohort performance alone is insufficient to judge whether a prediction model can be transported to another setting. The TRIPOD statement emphasizes that prediction model studies should transparently report the processes of model development, validation, and performance evaluation, including discrimination, calibration, and information related to clinical applicability.^23–25 In this study, Model 2C maintained good discrimination in the external validation cohort. However, good discrimination does not necessarily mean that predicted probabilities are adequately calibrated. External validation indicated a degree of risk overestimation, which may be related to differences between the two centers in patient composition, testing procedures, treatment management, PNR adjudication practices, or sample size. This finding also indicates the need for further evaluation of model calibration.

Accordingly, Model 2C is currently more suitable for relative risk stratification than for direct interpretation as an individual absolute risk estimate. The fixed cutoff of 0.2175 was derived from the derivation cohort and applied unchanged to the external validation cohort, avoiding optimism bias caused by threshold re-optimization in the validation cohort. This cutoff identified a low-risk group with a lower PNR rate in both cohorts. Its main value lies in early identification of patients at low risk of PNR, rather than serving as a basis for treatment intensification or treatment switching. For patients classified as high risk, the model result is better interpreted as a signal for enhanced early reassessment and comprehensive evaluation, which should still be integrated with symptoms, inflammatory markers, endoscopy, imaging, drug concentrations, anti-drug antibodies, and other relevant information.

The present data do not show that acting directly on this cutoff would improve patient outcomes. DCA can be used to evaluate the potential clinical net benefit of a prediction model across different threshold probabilities, but DCA findings are not equivalent to clinical impact studies. If model predictions are to be used to guide specific clinical actions, interpretation should be based on clearly defined clinical thresholds, specific implementation scenarios, and subsequent prospective validation or clinical impact studies.^26–28

Strengths, Limitations, and Future Directions

The strengths of this study include its two-center real-world cohort design, with Center 1 used for model development and Center 2 used for external validation, allowing preliminary assessment of the feasibility of cross-center application. Candidate predictors were restricted to routinely available pretreatment and Week 2 indicators, avoiding the inclusion of information from or after the outcome assessment time point and thereby reducing the risk of information leakage. The model included only three variables, facilitating clinical interpretation and future validation. In addition, this study evaluated model performance from multiple perspectives, including discrimination, calibration, overall prediction error, DCA, and fixed-cutoff risk stratification. The use of a nomogram and patient-level consistency checks also improved the transparency of model visualization.

This study also has limitations. Its retrospective design, variability in testing times, heterogeneity of clinical records, and retrospective reconstruction of CDAI may have affected variable extraction and outcome adjudication. The PNR definition combined clinical non-response with persistent objective inflammation at Week 14 and treatment failure events related to insufficient efficacy within Days 0–112. Although this approach reflects real-world clinical processes, it may also introduce heterogeneity in outcome adjudication. The limited sample size and number of PNR events restricted the stability of more complex model development, subgroup analyses, and additional sensitivity analyses. External validation was still preliminary and conducted within the healthcare system of the same region; therefore, the generalizability of the model requires further evaluation in broader regions and larger populations. In addition, the primary Model 2C analysis used complete-case analysis, although multiple imputation sensitivity analysis showed comparable results. Because the model depends on Week 2 CRP and albumin, it remains mainly applicable to patients who can complete Week 2 laboratory assessment. Residual bias related to the missingness mechanism cannot be completely excluded. IFX drug concentrations, anti-drug antibodies, fecal calprotectin, quantitative intestinal ultrasound parameters, more detailed endoscopic indicators, and molecular biomarkers were not included in the main model, mainly because of real-world accessibility and sample completeness considerations.

Future studies should further validate Model 2C in larger, multiregional, multicenter prospective cohorts and assess its stability, need for recalibration, and clinical impact across different healthcare systems and PNR definitions.

Conclusion

This study developed and performed preliminary external validation of an early risk stratification model for PNR in patients with CD treated with IFX, incorporating complicated disease behavior, Week 2 CRP, and Week 2 albumin. The main predictive contribution of the model was derived from Week 2 residual inflammatory burden and nutrition–inflammation status. Risk stratification using the fixed cutoff of 0.2175 derived using the Youden index in the derivation cohort may help identify patients at low risk of PNR. Given the risk overestimation observed during external validation, Model 2C should not be used as a standalone basis for treatment intensification or treatment switching. Further multicenter prospective validation is needed, together with assessment of model recalibration and clinical impact.

Funding Statement

This work was supported by the Henan Province Medical Science and Technology Research Plan Joint Construction Project (LHGJ20210419 and LHGJ20260408) and the Natural Science Foundation of Henan Province (242300421282).

Abbreviations

AUC, area under the receiver operating characteristic curve; BMI, body mass index; CD, Crohn’s disease; CDAI, Crohn’s Disease Activity Index; CDEIS, Crohn’s Disease Endoscopic Index of Severity; CI, confidence interval; CRP, C-reactive protein; CTE, computed tomography enterography; DCA, decision curve analysis; ECCO, European Crohn’s and Colitis Organisation; ESR, erythrocyte sedimentation rate; IFX, infliximab; IQR, interquartile range; LASSO, least absolute shrinkage and selection operator; MRE, magnetic resonance enterography; O/E, observed-to-expected; OR, odds ratio; PNR, primary non-response; ROC, receiver operating characteristic; SES-CD, Simple Endoscopic Score for Crohn’s Disease; T2T, treat-to-target; TDM, therapeutic drug monitoring; TNF, tumor necrosis factor; TRIPOD, Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis; ULN, upper limit of normal.

Data Sharing Statement

De-identified data underlying the findings of this study may be obtained from the corresponding author upon reasonable request.

Ethical Statement

Ethics approval was obtained from the ethics committees of the Second Affiliated Hospital of Zhengzhou University and the First Affiliated Hospital of Zhengzhou University (KY2025457 and 2026-KY-0285-002, respectively). The study complied with the Declaration of Helsinki. Because this was a retrospective observational study using de-identified data, informed consent was waived by the ethics committees.

Author Contributions

Meihua Zhao: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Visualization, Writing – original draft, Writing – review & editing. Dongyao Zhao: Conceptualization, Data curation, Investigation, Methodology, Validation, Writing – original draft, Writing – review & editing. Li Geng: Data curation, Investigation, Methodology, Validation, Writing – review & editing. Sumin Wang: Data curation, Investigation, Methodology, Validation, Writing – review & editing. Binbin Wang: Data curation, Investigation, Methodology, Validation, Writing – review & editing. Baisui Feng: Conceptualization, Funding acquisition, Project administration, Resources, Supervision, Writing – review & editing. All authors gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors report no conflicts of interest in this work.