Gradient Boosting Machine-Based Prognostic Model with SurvSHAP Analysis for Predicting Secondary Loss of Response to Infliximab Therapy in Crohn’s Disease Authors: Yu-Qing Guo, Yi-Ting Wang, Min Gao, Tao Liu, Li-Hong Huang, Sen Deng, Xiao-Yan Li, Meng-Dan Xu Published: 2026 ABSTRACT BackgroundInfliximab (IFX), a cornerstone anti-TNF-α therapy for moderate-to-severe Crohn’s disease (CD), is limited by secondary loss of response (SLOR), yet existing prognostic tools lack sufficient accuracy and interpretability. We developed and internally validated an interpretable machine learning (ML) framework for time-dependent SLOR prediction in IFX-treated CD patients.MethodsWe retrospectively enrolled 637 consecutive CD patients initiating IFX at a tertiary center. Multimodal predictors spanning demographic, pharmacokinetic (IFX trough concentrations and antibodies to infliximab [ATI] during induction and maintenance), hematological, biochemical, nutritional and coagulation domains were collected. Patients were randomly allocated to training (70%) and testing (30%) sets, stratified by SLOR status, and variables were screened via univariable and multivariable Cox regression. Nine ML algorithms were benchmarked at 12, 24 and 36 months using time-dependent AUC, concordance index, calibration and decision curve analysis. The top-performing model was further evaluated through risk-stratified Kaplan-Meier analysis and time-dependent SHapley Additive exPlanations (SurvSHAP).ResultsSeven independent determinants of SLOR were identified: maintenance-phase IFX trough concentration, disease duration, induction-phase ATI, platelet count, sex, concomitant immunosuppressant use and white blood cell count. The gradient boosting machine (GBM) demonstrated the strongest overall performance in the testing set, with time-dependent AUCs of 0.696, 0.825 and 0.854 at 12, 24 and 36 months, respectively, although discrimination at 12 months remained modest. GBM-based risk stratification yielded significantly divergent SLOR-free survival curves (HR = 6.462, 95% CI: 3.870–10.788, P < 0.001). SurvSHAP analysis ranked maintenance-phase IFX trough concentration as the predominant contributor with the widest effect range, followed by disease duration, induction-phase ATI and platelet count as adverse factors, while concomitant immunosuppressant use exhibited a protective signal.ConclusionThis internally validated ML framework showed improving discrimination over longer horizons, with maintenance-phase IFX trough concentration emerging as the dominant predictor, supporting proactive therapeutic drug monitoring pending external multicenter validation. Introduction Crohn’s disease (CD) is a chronic, transmural inflammatory disorder of the gastrointestinal tract associated with a rising global burden. Population-level analyses indicate that inflammatory bowel disease (IBD), of which CD constitutes a major phenotype, now affects millions worldwide, with accelerating incidence in newly industrialized regions and persistent prevalence growth in high-income countries, collectively driving progressive bowel damage, recurrent hospitalization, surgical intervention and substantially diminished health-related quality of life.1 Infliximab (IFX), a chimeric monoclonal antibody targeting soluble and membrane-bound tumor necrosis factor-alpha (TNF-α), occupies a central position within the biologic armamentarium for moderate-to-severe CD. By disrupting TNF-α-mediated inflammatory cascades, IFX promotes clinical remission and mucosal healing across induction and maintenance phases, with therapeutic drug monitoring (TDM) of IFX trough concentrations and antibodies to infliximab (ATI) further refining long-term dosing strategies.2,3 Notwithstanding these benefits, secondary loss of response (SLOR) remains a clinically consequential challenge. SLOR, manifesting as disease recurrence following a documented induction response, frequently necessitates regimen intensification or biologic switching. Prospective multicenter data from the PANTS cohort indicate that cumulative SLOR rates approach 60% by the third year of treatment among initial responders.4 The mechanistic underpinnings of SLOR are multifactorial and frequently involve overlapping pathways. Immunogenicity-driven ATI formation accelerates IFX clearance and neutralizes therapeutic activity,5 while pharmacokinetic variability, including heightened fecal drug loss and altered systemic clearance, contributes to subtherapeutic trough concentrations.6 Concurrently, disease-intrinsic factors encompassing progressive fibrostenotic remodeling, sustained mucosal inflammation and a shift toward non-TNF-α-driven immunological mechanisms further contribute to treatment failure.7 Against this backdrop, several predictive frameworks have sought to anticipate SLOR, yet persistent methodological shortcomings constrain their translational utility. Rong et al constructed a multimodal deep learning framework integrating baseline endoscopic imaging with clinical variables, achieving area under the curve (AUC) values of 0.824 to 0.892; however, heavy reliance on imaging data restricted external generalizability across heterogeneous populations.8 Qiu et al applied an extreme gradient boosting approach incorporating longitudinal laboratory trends, identifying hemoglobin, inflammatory markers, and Montreal classification as key predictors, though the retrospective design precluded comprehensive survival analysis and time-dependent interpretability.9 Feng et al developed a magnetic resonance imaging (MRI)-based radiomic signature for SLOR identification, yet modest sample sizes and omission of pharmacokinetic parameters limited broader applicability.10 Yang et al developed a CT-based pancreatic radiomics ML model to predict SLOR to IFX in biologic-naïve CD patients, achieving promising discriminatory performance; however, its reliance on single-modality baseline imaging without integration of pharmacokinetic or longitudinal variables limited its longitudinal applicability.11 These collective limitations highlight several critical research gaps. Most existing frameworks inadequately integrate high-dimensional multimodal data spanning clinical, pharmacokinetic, hematological and nutritional domains, and lack the capacity to capture nonlinear dose-response relationships or temporal risk trajectories. Conventional statistical approaches, including logistic regression and Cox proportional hazards models, remain further constrained by high-dimensional feature interactions.12 Advanced ML algorithms address these shortcomings by efficiently modeling nonlinear patterns across large, heterogeneous feature spaces while supporting clinical interpretability through time-dependent SHapley Additive exPlanations (SurvSHAP) analysis.13 The present study therefore sought to construct and internally validate a SurvSHAP-enhanced interpretable ML prognostic framework for SLOR prediction in CD patients receiving IFX therapy. Nine survival algorithms were systematically evaluated using comprehensive baseline and longitudinal variables, with nonlinear associations characterized via restricted cubic splines (RCS) and temporal interpretability achieved through SurvSHAP analysis. By enabling early individualized risk stratification, guiding TDM strategies and informing timely regimen modification, this framework is intended to support precision-medicine approaches that may reduce treatment failure and improve long-term outcomes in CD management.14 Materials and Methods Dataset and Target Population This retrospective cohort study was conducted at the Sixth Affiliated Hospital of Sun Yat-sen University, enrolling patients with CD who received IFX therapy between January 2019 and June 2023. Enrollment required fulfillment of all four inclusion criteria: (1) age between 18 and 75 years at treatment initiation; (2) CD diagnosis confirmed via multidimensional evaluation incorporating clinical manifestations, laboratory indices, endoscopic characteristics, radiological features and histopathological findings, consistent with authoritative diagnostic criteria;15,16 (3) documented completion of the standard IFX induction regimen, defined as at least three intravenous IFX infusions administered at weeks 0, 2 and 6, followed by the first maintenance infusion at week 14; (4) achievement of a clinical response at the end of the induction period (typically week 14), defined as a reduction in Crohn’s Disease Activity Index (CDAI) score of ≥ 70 points from baseline or a total CDAI ≤ 150.2,17 Patients were excluded if they fulfilled any of the following criteria: (1) failure to complete the induction phase or the week-14 response assessment; (2) pregnancy or lactation at any point during the observation period; (3) severe comorbidities substantially limiting follow-up or outcome adjudication, including end-stage hepatic dysfunction (Child-Pugh class C), end-stage renal disease (estimated glomerular filtration rate below 15 mL/min/1.73 m2) or malignancy; (4) known hypersensitivity to IFX or any of its excipients; (5) unavailability of key clinical data. Ethical approval was granted by the Institutional Review Board of the Sixth Affiliated Hospital of Sun Yat-sen University (No. 2023ZSLYEC-626). Given the retrospective, non-interventional design of this observational study, the requirement for individual informed consent was waived by the IRB. This study was conducted and reported in accordance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement to ensure methodological transparency and reproducibility of the prognostic model. Data Collection and Variables Clinical, pharmacokinetic and laboratory data were retrieved from electronic medical records at baseline and throughout the follow-up period. Disease-related characteristics documented at IFX initiation included sex, smoking history, alcohol consumption history, family history of CD, age at diagnosis, disease location per the Montreal Classification (ileal, colonic, ileocolonic or upper gastrointestinal), disease behavior (inflammatory, stricturing or penetrating), CDAI score, perianal disease status, prior intestinal or perianal surgical history, disease-related complications and extraintestinal manifestations (EIMs). Medication records captured prior and concomitant exposure to mesalazine, corticosteroids, immunosuppressants (IMM) and biologics. A series of continuous variables were additionally collected across pharmacokinetic, hematological, biochemical and nutritional domains. Pharmacokinetic parameters included IFX trough level during induction, antibodies to infliximab (ATI) level during induction, IFX trough concentration during maintenance, ATI level during maintenance and weight-adjusted IFX dose during both induction and maintenance phases. Patient-level indices comprised age at IFX initiation, disease duration and body mass index (BMI). Inflammatory markers included C-reactive protein (CRP) and erythrocyte sedimentation rate (ESR). Hematological indices encompassed white blood cell count (WBC), red blood cell count (RBC), hemoglobin (HGB), platelet count (PLT), absolute lymphocyte count (ALC) and absolute monocyte count (AMC). Hepatic biochemical parameters consisted of alanine aminotransferase (ALT), aspartate aminotransferase (AST), gamma-glutamyl transferase (GGT), alkaline phosphatase (ALP), and total bilirubin (TBIL). Metabolic and nutritional variables included total cholesterol (TC), creatinine (Cr), total protein (TP), albumin (ALB) and retinol-binding protein (RBP). Coagulation parameters comprised thrombin time (TT), fibrinogen (FIB), international normalized ratio (INR) and activated partial thromboplastin time (APTT). IFX Treatment Protocol and Therapeutic Drug Monitoring Induction therapy comprised intravenous IFX at 5 mg/kg administered at weeks 0, 2 and 6, followed by maintenance infusions at 8-week intervals. Serum trough samples were collected immediately prior to each scheduled infusion, with IFX trough levels and corresponding ATI titers concurrently quantified across both the induction and maintenance phases using a commercially validated enzyme-linked immunosorbent assay (ELISA; Immundiagnostik, Bensheim, Germany). The induction-phase IFX trough concentration was defined as the trough level measured immediately prior to the fourth IFX infusion (week 14). The maintenance-phase IFX trough concentration was defined as the most recent maintenance-phase trough level measured prior to the SLOR event for patients who developed SLOR, or the most recent maintenance-phase trough level prior to the last follow-up visit for event-free patients. Induction- and maintenance-phase ATI titers were defined at the corresponding time points. Definition of Clinical Outcomes The primary endpoint of this study was SLOR to IFX in patients with CD. Only patients who achieved a clinical response during the induction phase were eligible for the maintenance-phase analysis and were subsequently followed for the occurrence of SLOR. SLOR was defined as an inadequate clinical response during maintenance IFX therapy, evidenced by a CDAI score exceeding 150 points or a reduction in CDAI of less than 70 points, that necessitated at least one of the following treatment modifications: dose escalation of IFX, shortening of the IFX dosing interval, switching to an alternative biologic agent, or discontinuation of IFX due to treatment failure.10,18 Time to event was calculated from the date of the first IFX infusion to the date of SLOR confirmation. Patients who did not experience SLOR were censored at the date of last clinical contact or at the end of the predefined observation window, whichever occurred first. Data Handling and Feature Selection Approaches All preprocessing and analyses were performed in R (version 4.5.0; R Foundation for Statistical Computing, Vienna, Austria) and Python (version 3.10; Python Software Foundation, Wilmington, DE, USA). Missing data patterns across the full cohort were first visualized to guide subsequent handling (Supplementary Figure 1). To prevent information leakage, the cohort was partitioned into training (70%) and testing (30%) sets by stratified random sampling, preserving comparable event rates. Missing values were then imputed in a split-aware manner: the multiple imputation by chained equations (MICE) model was fitted on the training set alone and applied to the testing set. Multicollinearity was assessed within the training set using variance inflation factors (VIF) and tolerance values; variables with VIF > 10 or tolerance < 0.1 were excluded (Supplementary Table 1). Candidate variables were screened by univariable Cox proportional hazards regression, and those with P < 0.05 were entered into multivariable Cox regression to identify independent prognostic determinants of SLOR. Variables retaining significance (P < 0.05) constituted the final feature set for ML model construction. To characterize dose-response relationships between continuous predictors and SLOR risk, restricted cubic spline (RCS) regression was applied within the Cox framework on the full dataset. This descriptive analysis was independent of the ML pipeline; the non-linear associations were used solely for biological interpretation and were not included as ML model inputs. All predictive modeling steps, including feature screening, hyperparameter tuning and model fitting, were confined to the training set. For reproducibility, a fixed random seed of 123 was used across all stochastic procedures (data partitioning, MICE imputation, and model training); hyperparameter grids and final tuned values are listed in Supplementary Table 2, and software package versions in Supplementary Table 3. Model Development and Evaluation Nine ML algorithms were applied to the training set for time-to-SLOR prediction, encompassing gradient boosting machine (GBM), random survival forest (RSF), extreme gradient boosting (XGBoost), least absolute shrinkage and selection operator Cox regression (Lasso-Cox), Cox proportional hazards boosting (CoxBoost), partial least squares regression Cox model (plsRcox), deep survival neural network (Deepsurv), supervised principal component analysis (SuperPC), and support vector machine for survival (SurvivalSVM). Hyperparameter tuning for each algorithm was conducted through 10-fold cross-validation using grid search within the training set. Discriminative performance was assessed using time-dependent AUC and concordance index (C-index) at 12, 24 and 36 months. Calibration was examined through calibration curves and the Brier score, while clinical utility was quantified through decision curve analysis (DCA). Pairwise differences in time-dependent AUC were assessed using DeLong’s method, and multi-dimensional performance profiles across AUC, F1 score, accuracy, sensitivity and specificity were assessed through radar charts. Following identification of the optimal model, patients were dichotomized into low- and high-risk groups based on the median of the model-predicted risk scores.19 The median was pre-specified as a data-independent cut-off to avoid optimism from data-driven threshold selection and to yield two equally sized strata for the subsequent subgroup analyses. Kaplan-Meier survival curves were then constructed to compare SLOR-free survival probabilities between strata over the full 36-month follow-up period using the Log rank test. Hazard ratios (HR) with 95% confidence intervals (CI) for the high-risk versus low-risk group were estimated by univariable Cox proportional hazards regression. Model Interpretability Model interpretability was addressed through SurvSHAP for survival analysis, a framework that extends the game-theory-based SHAP approach to time-to-event outcomes by quantifying each feature’s contribution to predicted outcomes at individual observed time points.20 A survival explainer was built for the fitted GBM with a single MICE-imputed training set as background data. Individual survival functions were reconstructed as S(t | x) = exp{−H0(t)·exp(lp(x))}, where lp(x) was obtained from the GBM with the optimal number of trees, and H0(t) was estimated via the Breslow estimator from a Cox model of the observed outcome on lp(x) in the training set. SurvSHAP(t) values were computed for all testing set patients via model_survshap at the unique observed event times. Global feature importance was ranked based on the mean absolute SurvSHAP(t), aggregated via the integral method. To illustrate the dynamic nature of these effects, temporal trajectory plots of the average |SurvSHAP(t)| were generated, delineating how the influence of each predictor evolved throughout the follow-up period. Furthermore, patient-level time-aggregated SurvSHAP values were visualized using beeswarm plots, with color gradients denoting the min-max normalized feature magnitudes. Statistical Analysis Continuous variables were first examined for normality using the Shapiro–Wilk test; normally distributed data were expressed as mean ± standard deviation (SD) and compared by independent samples t-test, whereas non-normally distributed data were reported as median with interquartile range and analyzed by Mann–Whitney U-test. Categorical variables were presented as frequency and percentage and compared by Pearson chi-squared test, with Fisher’s exact test substituted when more than 20% of cells carried expected counts below five. Cox regression outputs were expressed as HR with 95% CI. P-values and HR estimates with 95% CI were reported to three decimal places, while all other descriptive measures were rounded to two decimal places. A two-tailed P < 0.05 was considered statistically significant. Results Patient Demographics and Cox Regression-Based Predictor Selection A total of 637 patients with CD receiving IFX were included, comprising 446 in the training set and 191 in the testing set. Overall, 331 patients developed SLOR (51.96%), including 240 (53.81%) in the training set and 91 (47.64%) in the testing set (Supplementary Figure 2). No statistically significant differences in baseline characteristics were detected between the training and testing sets (all P > 0.05; Table 1). Among the variables assessed in the univariable Cox analysis, 22 were significantly associated with SLOR (P < 0.05), including sex, age at diagnosis, behavior at diagnosis, prior intestinal surgery, complications, prior mesalazine use, concomitant IMM use, age at IFX initiation, disease duration, ATI level during induction, IFX trough concentration during maintenance, ATI level during maintenance, IFX dose during maintenance, BMI, ESR, CRP, WBC, ALC, HGB, PLT, Cr and ALB (Table 2).Table 1Baseline Demographic and Clinical Features Stratified by Training and Testing SetsVariablesTotalTraining Set (n=446)Testing Set (n=191)P valueSex, n (%)0.454Female481 (75.51)341 (76.46)140 (73.30)Male156 (24.49)105 (23.54)51 (26.70)Smoking history, n (%)0.542No607 (95.29)423 (94.84)184 (96.34)Yes30 (4.71)23 (5.16)7 (3.66)Alcohol drinking history, n (%)0.291No628 (98.59)438 (98.21)190 (99.48)Yes9 (1.41)8 (1.79)1 (0.52)Family history, n (%)0.186No631 (99.06)440 (98.65)191 (100.00)Yes6 (0.94)6 (1.35)0 (0.00)Age at diagnosis, n (%)0.415≤1638 (5.97)23 (5.16)15 (7.85)17–40560 (87.91)395 (88.57)165 (86.39)>4039 (6.12)28 (6.28)11 (5.76)Location at diagnosis, n (%)0.318Ileal67 (10.52)51 (11.43)16 (8.38)Colonic27 (4.24)21 (4.71)6 (3.14)Ileocolonic490 (76.92)341 (76.46)149 (78.01)Upper gastrointestinal53 (8.32)33 (7.40)20 (10.47)Behavior at diagnosis, n (%)0.415Inflammatory disease363 (56.99)247 (55.38)116 (60.73)Stricturing disease112 (17.58)83 (18.61)29 (15.18)Penetrating disease162 (25.43)116 (26.01)46 (24.08)CDAI, n (%)0.202Remission123 (19.31)79 (17.71)44 (23.04)Mild337 (52.90)234 (52.47)103 (53.93)Moderate156 (24.49)116 (26.01)40 (20.94)Severe21 (3.30)17 (3.81)4 (2.09)Perianal disease, n (%)0.937No227 (35.64)158 (35.43)69 (36.13)Yes410 (64.36)288 (64.57)122 (63.87)Prior intestinal surgery, n (%)0.305No454 (71.27%)312 (69.96%)142 (74.35%)Yes183 (28.73%)134 (30.04%)49 (25.65%)Prior perianal surgery, n (%)0.217No372 (58.40)268 (60.09)104 (54.45)Yes265 (41.60)178 (39.91)87 (45.55)Complications, n (%)0.505No319 (50.08)219 (49.10)100 (52.36)Yes318 (49.92)227 (50.90)91 (47.64)EIMs, n (%)0.906No527 (82.73)370 (82.96)157 (82.20)Yes110 (17.27)76 (17.04)34 (17.80)Prior mesalazine use, n (%)0.197No297 (46.62)200 (44.84)97 (50.79)Yes340 (53.38)246 (55.16)94 (49.21)Prior corticosteroid use, n (%)0.853No452 (70.96)315 (70.63)137 (71.73)Yes185 (29.04)131 (29.37)54 (28.27)Prior IMM use, n (%)1.000No401 (62.95)281 (63.00)120 (62.83)Yes236 (37.05)165 (37.00)71 (37.17)Prior biologic use, n (%)0.432No580 (91.05)403 (90.36%)177 (92.67)Yes57 (8.95)43 (9.64%)14 (7.33)Concomitant mesalazine use, n (%)0.456No611 (95.92)430 (96.41)181 (94.76)Yes26 (4.08)16 (3.59)10 (5.24)Concomitant corticosteroid use, n (%)0.497No627 (98.43)440 (98.65)187 (97.91)Yes10 (1.57)6 (1.35)4 (2.09)Concomitant IMM use, N (%)0.883No368 (57.77)259 (58.07)109 (57.07)Yes269 (42.23)187 (41.93)82 (42.93)Age at IFX initiation, years28.00 (23.00, 33.00)28.00 (23.00, 34.00)26.00 (22.00, 32.00)0.051Disease duration, years2.00 (1.00, 5.00)2.00 (1.00, 5.00)2.00 (1.00, 5.00)0.936IFX trough level during induction, mg/L4.68 (2.06, 12.32)4.75 (2.06, 11.82)4.64 (2.14, 13.32)0.796ATI level during induction, AU/mL3.74 (2.05, 7.34)3.67 (1.99, 7.19)3.94 (2.23, 7.83)0.492IFX trough concentration during maintenance, mg/L2.96 (0.86, 6.56)3.18 (0.90, 6.59)2.64 (0.78, 6.50)0.668ATI level during maintenance, AU/mL4.60 (2.91, 14.64)4.59 (2.83, 14.52)4.60 (3.00, 14.56)0.613IFX dose during induction, mg/kg5.68 (5.19, 6.15)5.69 (5.17, 6.15)5.68 (5.26, 6.25)0.881IFX dose during maintenance, mg/kg5.97 (5.45, 6.67)5.98 (5.45, 6.67)5.88 (5.38, 6.67)0.640BMI, kg/m219.17 (17.59, 20.99)19.13 (17.51, 20.96)19.33 (17.73, 21.06)0.709ESR, mm/h20.00 (9.00, 36.00)20.00 (9.00, 37.75)20.00 (10.00, 34.00)0.920CRP, mg/L4.80 (1.35, 17.33)4.55 (1.26, 17.65)5.49 (1.66, 16.73)0.572WBC, 109/L6.18 (4.75, 7.64)6.16 (4.62, 7.71)6.19 (5.05, 7.56)0.623ALC, 109/L1.35 (1.02, 1.71)1.32 (1.02, 1.69)1.38 (1.04, 1.75)0.239RBC, 1012/L4.63 (4.15, 5.06)4.63 (4.15, 5.07)4.65 (4.15, 4.96)0.961HGB, g/L125.00 (111.00, 138.00)124.90 (111.00, 137.00)127.00 (111.50, 139.00)0.503PLT, 109/L291.00 (236.00, 361.00)290.00 (232.25, 358.75)293.00 (245.00, 362.50)0.423AMC, 109/L0.51 (0.38, 0.66)0.50 (0.37, 0.65)0.52 (0.41, 0.68)0.186ALT, U/L14.24 (9.90, 22.05)14.41 (9.58, 22.31)13.55 (10.52, 18.89)0.701AST, U/L17.42 (14.03, 23.01)17.49 (14.05, 23.55)17.31 (13.90, 22.15)0.449GGT, U/L18.91 (13.96, 28.96)19.15 (14.04, 30.06)18.11 (13.71, 26.94)0.221ALP, U/L79.08 (65.63, 94.11)78.88 (65.83, 93.92)79.97 (65.12, 95.35)0.992TBIL, umol/L10.12 (7.78, 13.81)10.09 (7.67, 13.80)10.21 (8.04, 13.79)0.607TC, mmol/L3.64 (3.11, 4.21)3.64 (3.12, 4.21)3.61 (3.10, 4.15)0.907Cr, umol/L71.09 (59.73, 80.08)70.56 (59.73, 79.98)72.37 (60.84, 80.31)0.469TP, g/L70.67 ± 7.2770.74 ± 7.3070.52 ± 7.210.727ALB, g/L39.79 ± 5.1839.80 ± 5.1839.77 ± 5.190.952RBP, mg/L36.00 (28.50, 44.54)36.50 (28.52, 44.92)34.60 (28.39, 44.05)0.515INR1.08 (1.03, 1.15)1.09 (1.03, 1.16)1.08 (1.03, 1.14)0.744APTT, s31.80 (29.50, 34.30)31.50 (29.25, 34.27)32.40 (30.45, 34.53)0.053TT, s14.00 (13.10, 14.90)14.00 (13.10, 15.00)14.00 (13.20, 14.80)0.980FIB, g/L3.30 (2.70, 4.00)3.30 (2.64, 3.96)3.40 (2.80, 4.00)0.444Abbreviations: IFX, infliximab; CD, Crohn’s disease; SLOR, secondary loss of response; CDAI, Crohn’s disease activity index; EIMs, extraintestinal manifestations; IMM, immunosuppressant; ATI, antibodies to infliximab; BMI, body mass index; ESR, erythrocyte sedimentation rate; CRP, C-reactive protein; WBC, white blood cell count; ALC, absolute lymphocyte count; RBC, red blood cell count; HGB, hemoglobin; PLT, platelet count; AMC, absolute monocyte count; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, gamma-glutamyl transferase; ALP, alkaline phosphatase; TBIL, total bilirubin; TC, total cholesterol; Cr, creatinine; TP, total protein; ALB, albumin; RBP, retinol binding protein; INR, international normalized ratio; APTT, activated partial thromboplastin time; TT, thrombin time; FIB, fibrinogen. Table 2Independent Prognostic Determinants of SLOR to IFX in CD Identified by Univariable and Multivariable Cox Regression in the Training SetVariablesUnivariateP valueMultivariateP valueHR (95% CI)HR (95% CI)Sex, n (%)FemaleRefRefMale1.844 (1.398–2.433)˂0.0011.491 (1.048–2.120)0.026Smoking history, n (%)NoRefYes0.914 (0.511–1.633)0.761Alcohol drinking history, n (%)NoRefYes1.110 (0.458–2.694)0.817Family history, n (%)NoRefYes1.657 (0.616–4.456)0.317Age at diagnosis, n (%)≤16RefRef17–401.568 (0.831–2.957)0.1651.132 (0.551–2.323)0.736˃403.037 (1.444–6.386)0.0031.169 (0.417–3.277)0.766Location at diagnosis, n (%)IlealRefColonic0.883 (0.449–1.737)0.719Ileocolonic0.944 (0.634–1.406)0.777Upper gastrointestinal0.825 (0.452–1.508)0.532Behavior at diagnosis, n (%)Inflammatory diseaseRefRefStricturing disease0.883 (0.449–1.737)0.7191.024 (0.685–1.531)0.907Penetrating disease0.944 (0.634–1.406)0.7770.878 (0.578–1.333)0.542CDAI, n (%)RemissionRefMild0.825 (0.452–1.508)0.532Moderate1.290 (0.876–1.901)0.197Severe0.854 (0.383–1.900)0.698Perianal disease, n (%)NoRefYes0.780 (0.602–1.012)0.061Prior intestinal surgery, n (%)NoRefRefYes1.524 (1.171–1.985)0.0021.330 (0.895–1.975)0.158Prior perianal surgery, n (%)NoRefYes0.804 (0.618–1.046)0.104Complications, n (%)NoRefRefYes1.354 (1.049–1.746)0.0201.135 (0.806–1.598)0.470EIMs, N (%)NoRefYes1.217 (0.880–1.683)0.235Prior mesalazine use, n (%)NoRefRefYes1.546 (1.191–2.008)0.0011.177 (0.886–1.566)0.261Prior corticosteroid use, n (%)NoRefYes1.016 (0.769–1.342)0.911Prior IMM use, n (%)NoRefYes1.244 (0.961–1.610)0.098Prior biologic use, n (%)NoRefYes1.248 (0.842–1.851)0.270Concomitant mesalazine use, n (%)NoRefYes0.977 (0.460–2.072)0.951Concomitant corticosteroid use, n (%)NoRefYes1.728 (0.712–4.191)0.226Concomitant IMM use, n (%)NoRefRefYes0.662 (0.510–0.860)0.0020.757 (0.573–0.999)0.049Age at IFX initiation, years1.032 (1.018–1.047)˂0.0011.002 (0.979–1.026)0.857Disease duration, years1.039 (1.010–1.070)0.0091.059 (1.016–1.104)0.007IFX trough level during induction, mg/L1.000 (0.991–1.009)0.967ATI level during induction, AU/mL1.038 (1.026–1.050)˂0.0011.024 (1.012–1.037)˂0.001IFX trough concentration during maintenance, mg/L0.754 (0.713–0.798)˂0.0010.761 (0.715–0.810)˂0.001ATI level during maintenance, AU/mL1.003 (1.002–1.004)˂0.0011.000 (0.999–1.001)0.939IFX dose during induction, mg/kg1.086 (0.905–1.303)0.377IFX dose during maintenance, mg/kg1.250 (1.135–1.377)˂0.0011.060 (0.945–1.189)0.320BMI, kg/m20.946 (0.904–0.991)0.0180.986 (0.938–1.037)0.591ESR, mm/h1.006 (1.002–1.011)0.0100.999 (0.992–1.006)0.846CRP, mg/L1.006 (1.001–1.010)0.0111.000 (0.994–1.007)0.965WBC, 109/L1.018 (1.000–1.036)0.0471.022 (1.003–1.042)0.025ALC, 109/L0.738 (0.573–0.950)0.0180.770 (0.588–1.007)0.057AMC, 109/L0.880 (0.743–1.042)0.138RBC, 1012/L0.993 (0.988–0.999)0.022HGB, g/L1.002 (1.001–1.003)0.0011.006 (0.998–1.014)0.130PLT, 109/L1.314 (0.777–2.222)0.3081.003 (1.001–1.004)˂0.001ALT, U/L1.000 (0.996–1.005)0.853AST, U/L0.999 (0.993–1.005)0.738GGT, U/L1.000 (0.996–1.004)0.911ALP, U/L0.998 (0.994–1.003)0.438TBIL, umol/L0.983 (0.957–1.010)0.221TC, mmol/L0.996 (0.859–1.156)0.963Cr, umol/L0.989 (0.981–0.998)0.0151.000 (0.994–1.005)0.929TP, g/L0.993 (0.975–1.011)0.449ALB, g/L0.956 (0.932–0.980)˂0.0010.996 (0.965–1.027)0.786RBP, mg/L0.996 (0.986–1.007)0.525INR2.793 (0.927–8.420)0.068APTT, s0.991 (0.960–1.023)0.561TT, s1.000 (0.982–1.017)0.967FIB, g/L1.102 (0.965–1.260)0.152Abbreviations: IFX, infliximab; CD, Crohn’s disease; SLOR, secondary loss of response; Ref, reference; CDAI, Crohn’s disease activity index; EIMs, extraintestinal manifestations; IMM, immunosuppressant; ATI, antibodies to infliximab; BMI, body mass index; ESR, erythrocyte sedimentation rate; CRP, C-reactive protein; WBC, white blood cell count; ALC, absolute lymphocyte count; AMC, absolute monocyte count; RBC, red blood cell count; HGB, hemoglobin; PLT, platelet count; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, gamma-glutamyl transferase; ALP, alkaline phosphatase; TBIL, total bilirubin; TC, total cholesterol; Cr, creatinine; TP, total protein; ALB, albumin; RBP, retinol binding protein; INR, international normalized ratio; APTT, activated partial thromboplastin time; TT, thrombin time; FIB, fibrinogen. Table 1Baseline Demographic and Clinical Features Stratified by Training and Testing SetsVariablesTotalTraining Set (n=446)Testing Set (n=191)P valueSex, n (%)0.454Female481 (75.51)341 (76.46)140 (73.30)Male156 (24.49)105 (23.54)51 (26.70)Smoking history, n (%)0.542No607 (95.29)423 (94.84)184 (96.34)Yes30 (4.71)23 (5.16)7 (3.66)Alcohol drinking history, n (%)0.291No628 (98.59)438 (98.21)190 (99.48)Yes9 (1.41)8 (1.79)1 (0.52)Family history, n (%)0.186No631 (99.06)440 (98.65)191 (100.00)Yes6 (0.94)6 (1.35)0 (0.00)Age at diagnosis, n (%)0.415≤1638 (5.97)23 (5.16)15 (7.85)17–40560 (87.91)395 (88.57)165 (86.39)>4039 (6.12)28 (6.28)11 (5.76)Location at diagnosis, n (%)0.318Ileal67 (10.52)51 (11.43)16 (8.38)Colonic27 (4.24)21 (4.71)6 (3.14)Ileocolonic490 (76.92)341 (76.46)149 (78.01)Upper gastrointestinal53 (8.32)33 (7.40)20 (10.47)Behavior at diagnosis, n (%)0.415Inflammatory disease363 (56.99)247 (55.38)116 (60.73)Stricturing disease112 (17.58)83 (18.61)29 (15.18)Penetrating disease162 (25.43)116 (26.01)46 (24.08)CDAI, n (%)0.202Remission123 (19.31)79 (17.71)44 (23.04)Mild337 (52.90)234 (52.47)103 (53.93)Moderate156 (24.49)116 (26.01)40 (20.94)Severe21 (3.30)17 (3.81)4 (2.09)Perianal disease, n (%)0.937No227 (35.64)158 (35.43)69 (36.13)Yes410 (64.36)288 (64.57)122 (63.87)Prior intestinal surgery, n (%)0.305No454 (71.27%)312 (69.96%)142 (74.35%)Yes183 (28.73%)134 (30.04%)49 (25.65%)Prior perianal surgery, n (%)0.217No372 (58.40)268 (60.09)104 (54.45)Yes265 (41.60)178 (39.91)87 (45.55)Complications, n (%)0.505No319 (50.08)219 (49.10)100 (52.36)Yes318 (49.92)227 (50.90)91 (47.64)EIMs, n (%)0.906No527 (82.73)370 (82.96)157 (82.20)Yes110 (17.27)76 (17.04)34 (17.80)Prior mesalazine use, n (%)0.197No297 (46.62)200 (44.84)97 (50.79)Yes340 (53.38)246 (55.16)94 (49.21)Prior corticosteroid use, n (%)0.853No452 (70.96)315 (70.63)137 (71.73)Yes185 (29.04)131 (29.37)54 (28.27)Prior IMM use, n (%)1.000No401 (62.95)281 (63.00)120 (62.83)Yes236 (37.05)165 (37.00)71 (37.17)Prior biologic use, n (%)0.432No580 (91.05)403 (90.36%)177 (92.67)Yes57 (8.95)43 (9.64%)14 (7.33)Concomitant mesalazine use, n (%)0.456No611 (95.92)430 (96.41)181 (94.76)Yes26 (4.08)16 (3.59)10 (5.24)Concomitant corticosteroid use, n (%)0.497No627 (98.43)440 (98.65)187 (97.91)Yes10 (1.57)6 (1.35)4 (2.09)Concomitant IMM use, N (%)0.883No368 (57.77)259 (58.07)109 (57.07)Yes269 (42.23)187 (41.93)82 (42.93)Age at IFX initiation, years28.00 (23.00, 33.00)28.00 (23.00, 34.00)26.00 (22.00, 32.00)0.051Disease duration, years2.00 (1.00, 5.00)2.00 (1.00, 5.00)2.00 (1.00, 5.00)0.936IFX trough level during induction, mg/L4.68 (2.06, 12.32)4.75 (2.06, 11.82)4.64 (2.14, 13.32)0.796ATI level during induction, AU/mL3.74 (2.05, 7.34)3.67 (1.99, 7.19)3.94 (2.23, 7.83)0.492IFX trough concentration during maintenance, mg/L2.96 (0.86, 6.56)3.18 (0.90, 6.59)2.64 (0.78, 6.50)0.668ATI level during maintenance, AU/mL4.60 (2.91, 14.64)4.59 (2.83, 14.52)4.60 (3.00, 14.56)0.613IFX dose during induction, mg/kg5.68 (5.19, 6.15)5.69 (5.17, 6.15)5.68 (5.26, 6.25)0.881IFX dose during maintenance, mg/kg5.97 (5.45, 6.67)5.98 (5.45, 6.67)5.88 (5.38, 6.67)0.640BMI, kg/m219.17 (17.59, 20.99)19.13 (17.51, 20.96)19.33 (17.73, 21.06)0.709ESR, mm/h20.00 (9.00, 36.00)20.00 (9.00, 37.75)20.00 (10.00, 34.00)0.920CRP, mg/L4.80 (1.35, 17.33)4.55 (1.26, 17.65)5.49 (1.66, 16.73)0.572WBC, 109/L6.18 (4.75, 7.64)6.16 (4.62, 7.71)6.19 (5.05, 7.56)0.623ALC, 109/L1.35 (1.02, 1.71)1.32 (1.02, 1.69)1.38 (1.04, 1.75)0.239RBC, 1012/L4.63 (4.15, 5.06)4.63 (4.15, 5.07)4.65 (4.15, 4.96)0.961HGB, g/L125.00 (111.00, 138.00)124.90 (111.00, 137.00)127.00 (111.50, 139.00)0.503PLT, 109/L291.00 (236.00, 361.00)290.00 (232.25, 358.75)293.00 (245.00, 362.50)0.423AMC, 109/L0.51 (0.38, 0.66)0.50 (0.37, 0.65)0.52 (0.41, 0.68)0.186ALT, U/L14.24 (9.90, 22.05)14.41 (9.58, 22.31)13.55 (10.52, 18.89)0.701AST, U/L17.42 (14.03, 23.01)17.49 (14.05, 23.55)17.31 (13.90, 22.15)0.449GGT, U/L18.91 (13.96, 28.96)19.15 (14.04, 30.06)18.11 (13.71, 26.94)0.221ALP, U/L79.08 (65.63, 94.11)78.88 (65.83, 93.92)79.97 (65.12, 95.35)0.992TBIL, umol/L10.12 (7.78, 13.81)10.09 (7.67, 13.80)10.21 (8.04, 13.79)0.607TC, mmol/L3.64 (3.11, 4.21)3.64 (3.12, 4.21)3.61 (3.10, 4.15)0.907Cr, umol/L71.09 (59.73, 80.08)70.56 (59.73, 79.98)72.37 (60.84, 80.31)0.469TP, g/L70.67 ± 7.2770.74 ± 7.3070.52 ± 7.210.727ALB, g/L39.79 ± 5.1839.80 ± 5.1839.77 ± 5.190.952RBP, mg/L36.00 (28.50, 44.54)36.50 (28.52, 44.92)34.60 (28.39, 44.05)0.515INR1.08 (1.03, 1.15)1.09 (1.03, 1.16)1.08 (1.03, 1.14)0.744APTT, s31.80 (29.50, 34.30)31.50 (29.25, 34.27)32.40 (30.45, 34.53)0.053TT, s14.00 (13.10, 14.90)14.00 (13.10, 15.00)14.00 (13.20, 14.80)0.980FIB, g/L3.30 (2.70, 4.00)3.30 (2.64, 3.96)3.40 (2.80, 4.00)0.444Abbreviations: IFX, infliximab; CD, Crohn’s disease; SLOR, secondary loss of response; CDAI, Crohn’s disease activity index; EIMs, extraintestinal manifestations; IMM, immunosuppressant; ATI, antibodies to infliximab; BMI, body mass index; ESR, erythrocyte sedimentation rate; CRP, C-reactive protein; WBC, white blood cell count; ALC, absolute lymphocyte count; RBC, red blood cell count; HGB, hemoglobin; PLT, platelet count; AMC, absolute monocyte count; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, gamma-glutamyl transferase; ALP, alkaline phosphatase; TBIL, total bilirubin; TC, total cholesterol; Cr, creatinine; TP, total protein; ALB, albumin; RBP, retinol binding protein; INR, international normalized ratio; APTT, activated partial thromboplastin time; TT, thrombin time; FIB, fibrinogen. Baseline Demographic and Clinical Features Stratified by Training and Testing Sets Abbreviations: IFX, infliximab; CD, Crohn’s disease; SLOR, secondary loss of response; CDAI, Crohn’s disease activity index; EIMs, extraintestinal manifestations; IMM, immunosuppressant; ATI, antibodies to infliximab; BMI, body mass index; ESR, erythrocyte sedimentation rate; CRP, C-reactive protein; WBC, white blood cell count; ALC, absolute lymphocyte count; RBC, red blood cell count; HGB, hemoglobin; PLT, platelet count; AMC, absolute monocyte count; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, gamma-glutamyl transferase; ALP, alkaline phosphatase; TBIL, total bilirubin; TC, total cholesterol; Cr, creatinine; TP, total protein; ALB, albumin; RBP, retinol binding protein; INR, international normalized ratio; APTT, activated partial thromboplastin time; TT, thrombin time; FIB, fibrinogen. Table 2Independent Prognostic Determinants of SLOR to IFX in CD Identified by Univariable and Multivariable Cox Regression in the Training SetVariablesUnivariateP valueMultivariateP valueHR (95% CI)HR (95% CI)Sex, n (%)FemaleRefRefMale1.844 (1.398–2.433)˂0.0011.491 (1.048–2.120)0.026Smoking history, n (%)NoRefYes0.914 (0.511–1.633)0.761Alcohol drinking history, n (%)NoRefYes1.110 (0.458–2.694)0.817Family history, n (%)NoRefYes1.657 (0.616–4.456)0.317Age at diagnosis, n (%)≤16RefRef17–401.568 (0.831–2.957)0.1651.132 (0.551–2.323)0.736˃403.037 (1.444–6.386)0.0031.169 (0.417–3.277)0.766Location at diagnosis, n (%)IlealRefColonic0.883 (0.449–1.737)0.719Ileocolonic0.944 (0.634–1.406)0.777Upper gastrointestinal0.825 (0.452–1.508)0.532Behavior at diagnosis, n (%)Inflammatory diseaseRefRefStricturing disease0.883 (0.449–1.737)0.7191.024 (0.685–1.531)0.907Penetrating disease0.944 (0.634–1.406)0.7770.878 (0.578–1.333)0.542CDAI, n (%)RemissionRefMild0.825 (0.452–1.508)0.532Moderate1.290 (0.876–1.901)0.197Severe0.854 (0.383–1.900)0.698Perianal disease, n (%)NoRefYes0.780 (0.602–1.012)0.061Prior intestinal surgery, n (%)NoRefRefYes1.524 (1.171–1.985)0.0021.330 (0.895–1.975)0.158Prior perianal surgery, n (%)NoRefYes0.804 (0.618–1.046)0.104Complications, n (%)NoRefRefYes1.354 (1.049–1.746)0.0201.135 (0.806–1.598)0.470EIMs, N (%)NoRefYes1.217 (0.880–1.683)0.235Prior mesalazine use, n (%)NoRefRefYes1.546 (1.191–2.008)0.0011.177 (0.886–1.566)0.261Prior corticosteroid use, n (%)NoRefYes1.016 (0.769–1.342)0.911Prior IMM use, n (%)NoRefYes1.244 (0.961–1.610)0.098Prior biologic use, n (%)NoRefYes1.248 (0.842–1.851)0.270Concomitant mesalazine use, n (%)NoRefYes0.977 (0.460–2.072)0.951Concomitant corticosteroid use, n (%)NoRefYes1.728 (0.712–4.191)0.226Concomitant IMM use, n (%)NoRefRefYes0.662 (0.510–0.860)0.0020.757 (0.573–0.999)0.049Age at IFX initiation, years1.032 (1.018–1.047)˂0.0011.002 (0.979–1.026)0.857Disease duration, years1.039 (1.010–1.070)0.0091.059 (1.016–1.104)0.007IFX trough level during induction, mg/L1.000 (0.991–1.009)0.967ATI level during induction, AU/mL1.038 (1.026–1.050)˂0.0011.024 (1.012–1.037)˂0.001IFX trough concentration during maintenance, mg/L0.754 (0.713–0.798)˂0.0010.761 (0.715–0.810)˂0.001ATI level during maintenance, AU/mL1.003 (1.002–1.004)˂0.0011.000 (0.999–1.001)0.939IFX dose during induction, mg/kg1.086 (0.905–1.303)0.377IFX dose during maintenance, mg/kg1.250 (1.135–1.377)˂0.0011.060 (0.945–1.189)0.320BMI, kg/m20.946 (0.904–0.991)0.0180.986 (0.938–1.037)0.591ESR, mm/h1.006 (1.002–1.011)0.0100.999 (0.992–1.006)0.846CRP, mg/L1.006 (1.001–1.010)0.0111.000 (0.994–1.007)0.965WBC, 109/L1.018 (1.000–1.036)0.0471.022 (1.003–1.042)0.025ALC, 109/L0.738 (0.573–0.950)0.0180.770 (0.588–1.007)0.057AMC, 109/L0.880 (0.743–1.042)0.138RBC, 1012/L0.993 (0.988–0.999)0.022HGB, g/L1.002 (1.001–1.003)0.0011.006 (0.998–1.014)0.130PLT, 109/L1.314 (0.777–2.222)0.3081.003 (1.001–1.004)˂0.001ALT, U/L1.000 (0.996–1.005)0.853AST, U/L0.999 (0.993–1.005)0.738GGT, U/L1.000 (0.996–1.004)0.911ALP, U/L0.998 (0.994–1.003)0.438TBIL, umol/L0.983 (0.957–1.010)0.221TC, mmol/L0.996 (0.859–1.156)0.963Cr, umol/L0.989 (0.981–0.998)0.0151.000 (0.994–1.005)0.929TP, g/L0.993 (0.975–1.011)0.449ALB, g/L0.956 (0.932–0.980)˂0.0010.996 (0.965–1.027)0.786RBP, mg/L0.996 (0.986–1.007)0.525INR2.793 (0.927–8.420)0.068APTT, s0.991 (0.960–1.023)0.561TT, s1.000 (0.982–1.017)0.967FIB, g/L1.102 (0.965–1.260)0.152Abbreviations: IFX, infliximab; CD, Crohn’s disease; SLOR, secondary loss of response; Ref, reference; CDAI, Crohn’s disease activity index; EIMs, extraintestinal manifestations; IMM, immunosuppressant; ATI, antibodies to infliximab; BMI, body mass index; ESR, erythrocyte sedimentation rate; CRP, C-reactive protein; WBC, white blood cell count; ALC, absolute lymphocyte count; AMC, absolute monocyte count; RBC, red blood cell count; HGB, hemoglobin; PLT, platelet count; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, gamma-glutamyl transferase; ALP, alkaline phosphatase; TBIL, total bilirubin; TC, total cholesterol; Cr, creatinine; TP, total protein; ALB, albumin; RBP, retinol binding protein; INR, international normalized ratio; APTT, activated partial thromboplastin time; TT, thrombin time; FIB, fibrinogen. Independent Prognostic Determinants of SLOR to IFX in CD Identified by Univariable and Multivariable Cox Regression in the Training Set Abbreviations: IFX, infliximab; CD, Crohn’s disease; SLOR, secondary loss of response; Ref, reference; CDAI, Crohn’s disease activity index; EIMs, extraintestinal manifestations; IMM, immunosuppressant; ATI, antibodies to infliximab; BMI, body mass index; ESR, erythrocyte sedimentation rate; CRP, C-reactive protein; WBC, white blood cell count; ALC, absolute lymphocyte count; AMC, absolute monocyte count; RBC, red blood cell count; HGB, hemoglobin; PLT, platelet count; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, gamma-glutamyl transferase; ALP, alkaline phosphatase; TBIL, total bilirubin; TC, total cholesterol; Cr, creatinine; TP, total protein; ALB, albumin; RBP, retinol binding protein; INR, international normalized ratio; APTT, activated partial thromboplastin time; TT, thrombin time; FIB, fibrinogen. In the multivariable Cox analysis, seven variables remained independent predictors of SLOR. Male sex, disease duration, ATI level during induction, WBC and PLT were associated with an increased risk of SLOR, whereas concomitant IMM use and IFX trough concentration during maintenance were associated with a reduced risk (all P < 0.05; Table 2). Overall, these findings indicate that clinical, pharmacokinetic and immunogenic factors jointly contribute to the risk of SLOR. Notably, the training cohort provided 240 SLOR events for the seven independent predictors retained in the multivariable Cox model, corresponding to approximately 34 events per predictor, well above the widely accepted events-per-variable (EPV) ≥ 10 threshold, indicating that the sample size was adequate to support stable effect estimation and to minimize the risk of overfitting. Nonlinear Dose-Response Relationships Between Continuous Predictors and SLOR Risk To characterize the functional dose-response relationships between continuous clinical variables and SLOR risk, RCS analyses were incorporated into a Cox proportional hazards framework (Figure 1). Among the variables examined, disease duration showed a significant positive association with SLOR risk (P for overall = 0.003; P for nonlinear = 0.087), with estimated HRs increasing in an approximately linear fashion beyond the reference of 2 years. ATI levels during induction likewise reached statistical significance (P for overall < 0.001; P for nonlinear = 0.045), with risk escalating in a predominantly dose-dependent but nonlinear fashion across the observed titer spectrum relative to a reference of 3.74 AU/mL. In contrast to these positive associations, IFX trough concentrations during the maintenance phase revealed a pronounced nonlinear inverse relationship with SLOR risk (P for overall < 0.001; P for nonlinear < 0.001), with estimated HRs declining steeply once trough levels surpassed the reference threshold of 2.96 mg/L. Regarding hematological parameters, WBC did not attain statistical significance (P for overall = 0.159), whereas PLT constituted an independent predictor of SLOR (P for overall = 0.006, P for nonlinear = 0.195), with failure risk ascending monotonically beyond the reference of 291.00 × 109/L. Collectively, these RCS-derived dose-response trajectories underscore the heterogeneous contributions of individual continuous predictors, providing a biologically informed rationale for their subsequent incorporation into the ML prognostic framework. Figure 1Restricted cubic spline analyses depicting the relationships between continuous predictors and the risk of SLOR to IFX in patients with CD. (A) Disease duration; (B) ATI level during induction; (C) IFX trough concentration during maintenance; (D) WBC; (E) PLT.The image A showing a line graph with y-axis label, Hazard Ratio (95 percent CI), ranging from 0 to 8 and x-axis label, Disease duration, ranging from 0 to 25. Text: P for overall equals 0.003; P for nonlinear equals 0.087. A vertical dashed line at x equals 2.00 labeled Ref equals 2.00. A horizontal dashed line at y equals 1. The curve is near y equals 1 around x equals 0 to 5, then rises to about y equals 2.7 at x equals 25, with a shaded confidence band widening toward x equals 25. The image B showing a line graph with y-axis label, Hazard Ratio (95 percent CI), ranging from 0 to 10 and x-axis label, ATI level during induction, ranging from 0 to 80. Text: P for overall less than 0.001; P for nonlinear equals 0.045. A vertical dashed line at x equals 3.74 labeled Ref equals 3.74. A horizontal dashed line at y equals 1. The curve increases from near y equals 1 at low x to about y equals 3.6 at x equals 80, with a widening shaded band. The image C showing a line graph with y-axis label, Hazard Ratio (95 percent CI), ranging from 0 to 6 and x-axis label, IFX trough concentration during maintenance, ranging from 0 to 60. Text: P for overall less than 0.001; P for nonlinear less than 0.001. A vertical dashed line at x equals 2.96 labeled Ref equals 2.96. A horizontal dashed line at y equals 1. The curve starts near y equals 5 at x near 0, drops steeply below y equals 1 by about x equals 5 and approaches near y equals 0 by x equals 60; the shaded band is narrow midrange and widens slightly at higher x. The image D showing a line graph with y-axis label, Hazard Ratio (95 percent CI), ranging from 0 to 6 and x-axis label, WBC, ranging from 0 to 20. Text: P for overall equals 0.159; P for nonlinear equals 0.440. A vertical dashed line at x equals 6.18 labeled Ref equals 6.18. A horizontal dashed line at y equals 1. The curve declines from about y equals 2 at x equals 0 to near y equals 1 around x equals 6, then slowly rises to about y equals 1.4 at x equals 20; the shaded band is widest near x equals 0. The image E showing a line graph with y-axis label, Hazard Ratio (95 percent CI), ranging from 0 to 8 and x-axis label, PLT, ranging from 100 to 700. Text: P for overall equals 0.006; P for nonlinear equals 0.195. A vertical dashed line at x equals 291.00 labeled Ref equals 291.00. A horizontal dashed line at y equals 1. The curve is near y equals 1 around x equals 291, then rises to about y equals 2.8 at x equals 700; the shaded band widens markedly above about x equals 500.A set of five line graphs showing hazard ratio versus disease duration, ATI, IFX, WBC and PLT.Abbreviations: IFX, infliximab; SLOR, secondary loss of response; CD, Crohn’s disease; ATI, antibodies to infliximab; WBC, white blood cell count; PLT, platelet count. Restricted cubic spline analyses depicting the relationships between continuous predictors and the risk of SLOR to IFX in patients with CD. (A) Disease duration; (B) ATI level during induction; (C) IFX trough concentration during maintenance; (D) WBC; (E) PLT. Model-Based Comprehensive Predictive Performance Evaluation Nine ML algorithms were systematically evaluated through time-dependent AUC and C-index trajectories, calibration assessment, DCA and pairwise DeLong testing at 12, 24 and 36 months across training and testing sets to identify the optimal prognostic model for SLOR to IFX in CD patients (Figures 2 and 3; Supplementary Figures 3 and 4). Figure 2Time-dependent discriminative performance of nine machine learning models across training and testing sets. (A) Time-dependent AUC curves in the training set; (B) C-index curves in the training set; (C) Time-dependent AUC curves in the testing set; (D) C-index curves in the testing set.Image A: Line graph with Time (months) on x-axis and AUC on y-axis (0.00 to 1.00). Legend: Lasso Cox, CoxBoost, RSF, GBM, SurvivalSVM, Deepsurv, Xgboost, SuperPC, plsRcox. A dashed line at 0.70 AUC. Curves form two bands: upper band rises from 0.70 to 0.95 by 36 months, lower band stays around 0.50 to 0.60. Image B: Line graph with Time (months) on x-axis and C-index on y-axis (0.00 to 1.00). Same legend. Dashed line at 0.70 C-index. Upper band around 0.70 to 0.85, lower band around 0.50 to 0.60. Image C: Line graph with Time (months) on x-axis and AUC on y-axis (0.00 to 1.00). Same legend. Dashed line at 0.70 AUC. Curves peak near 0.85 early, dip to 0.70, rise to 0.85 by 36 months. Two curves stay lower, around 0.35 to 0.55. Image D: Line graph with Time (months) on x-axis and C-index on y-axis (0.00 to 1.00). Same legend. Dashed line at 0.70 C-index. Curves cluster around 0.70 to 0.78 after early fluctuations, two curves remain lower around 0.40 to 0.55.A four line graph set showing time dependent AUC and C-index for nine machine learning models.Abbreviations: AUC, area under the receiver operating characteristic curve; RSF, random survival forest; GBM, gradient boosting machine; SVM, support vector machine; XGBoost, extreme gradient boosting; SuperPC, supervised principal component; plsRcox, partial least squares regression Cox model; Lasso-Cox, least absolute shrinkage and selection operator Cox regression; CoxBoost, Cox proportional hazards boosting; Deepsurv, deep survival neural network. Figure 3Predictive performance of nine machine learning models assessed at 12, 24 and 36 months in the training and testing sets of CD patients at risk of SLOR to IFX. (A) ROC curves from the training set at 12 months. (B) ROC curves from the testing set at 12 months. (C) ROC curves from the training set at 24 months. (D) ROC curves from the testing set at 24 months. (E) ROC curves from the training set at 36 months. (F) ROC curves from the testing set at 36 months.Six graphs evaluate nine machine learning models using ROC curves, with axes for 1-specificity and sensitivity (0.00 to 1.00). Models: RSF, GBM, Lasso Cox, CoxBoost, SurvivalSVM, Deepsurv, Xgboost, SuperPC, plsRcox. Image A: RSF (AUC 0.884) and Xgboost (AUC 0.841) excel; SurvivalSVM (AUC 0.523) and SuperPC (AUC 0.529) lag. Image B: GBM (AUC 0.696) leads; SurvivalSVM (AUC 0.577) and SuperPC (AUC 0.581) trail. Image C: RSF (AUC 0.947) and Xgboost (AUC 0.896) top; SurvivalSVM (AUC 0.584) and SuperPC (AUC 0.593) bottom. Image D: Xgboost (AUC 0.807) highest; SurvivalSVM (AUC 0.496) and SuperPC (AUC 0.513) lowest. Image E: RSF (AUC 0.967) and Xgboost (AUC 0.944) excel; SurvivalSVM (AUC 0.554) and SuperPC (AUC 0.566) underperform. Image F: GBM (AUC 0.845) leads; SurvivalSVM (AUC 0.516) and SuperPC (AUC 0.518) lag. Overall, RSF and Xgboost consistently achieve high AUCs, while SurvivalSVM and SuperPC score low.A multi line graph set of 6 receiver operating characteristic plots A to F comparing nine models.Abbreviations: ROC, receiver operating characteristic; CD, Crohn’s disease; SLOR, secondary loss of response; IFX, infliximab; RSF, random survival forest; GBM, gradient boosting machine; SVM, support vector machine; XGBoost, extreme gradient boosting; SuperPC, supervised principal component; plsRcox, partial least squares regression Cox model; Lasso-Cox, least absolute shrinkage and selection operator Cox regression; CoxBoost, Cox proportional hazards boosting; Deepsurv, deep survival neural network. Time-dependent discriminative performance of nine machine learning models across training and testing sets. (A) Time-dependent AUC curves in the training set; (B) C-index curves in the training set; (C) Time-dependent AUC curves in the testing set; (D) C-index curves in the testing set. Predictive performance of nine machine learning models assessed at 12, 24 and 36 months in the training and testing sets of CD patients at risk of SLOR to IFX. (A) ROC curves from the training set at 12 months. (B) ROC curves from the testing set at 12 months. (C) ROC curves from the training set at 24 months. (D) ROC curves from the testing set at 24 months. (E) ROC curves from the training set at 36 months. (F) ROC curves from the testing set at 36 months. Time-dependent AUC and C-index curves revealed consistent discrimination patterns over the full follow-up period (Figure 2). In the training set, ensemble-based algorithms collectively stabilized above the 0.70 reference threshold after approximately 12 months, with RSF attaining the highest sustained values in both metrics, while SurvivalSVM remained below the threshold throughout (Figure 2A and B). In the testing set, the inter-model spread narrowed markedly, with most ensemble algorithms converging above 0.70, whereas SuperPC and SurvivalSVM failed to exceed this threshold at any timepoint (Figure 2C and D). Notably, GBM maintained a stable and well-sustained trajectory in the testing set, in contrast to the more variable patterns observed for several competing models, suggesting a possible generalizability advantage. At the three prespecified timepoints (12, 24 and 36 months), RSF achieved the highest training-set AUCs (0.884, 0.947 and 0.967, respectively), followed by XGBoost (0.841, 0.896 and 0.944, respectively) and GBM (0.790, 0.850 and 0.909, respectively), while SurvivalSVM and SuperPC exhibited consistently poor discrimination (AUC range: 0.523–0.593; Figure 3A, C and D). Within-sample performance alone, however, does not reliably reflect model generalizability, particularly for high-complexity algorithms prone to overfitting. In the testing set, the discriminative hierarchy was substantially reconfigured. GBM achieved the highest AUCs across all three timepoints (0.696, 0.825 and 0.854, respectively), consistently surpassing RSF (0.680, 0.813 and 0.814, respectively), while XGBoost (0.708, 0.818 and 0.823, respectively), CoxBoost (0.694, 0.798 and 0.831, respectively), Lasso-Cox (0.673, 0.803 and 0.828, respectively) and plsRcox (0.676, 0.793 and 0.818, respectively) maintained acceptable though comparatively lower values (Figure 3B, D and F). The modest AUC attenuation from the training to the testing set observed for GBM, relative to the more pronounced decline seen with RSF, is consistent with the trajectory-level observations and indicates a lower tendency to overfit and more stable predictive performance upon internal validation. Calibration assessment further differentiated model reliability (Supplementary Figure 3). In the training set, GBM, RSF, CoxBoost and plsRcox generated curves closely approximating the ideal diagonal, while SurvivalSVM exhibited pronounced systematic miscalibration with predicted probabilities confined to a narrow band. In the testing set, GBM maintained comparatively reasonable calibration at all three timepoints, particularly at 24 and 36 months, whereas RSF and several other algorithms showed more evident deviation from the reference line, supporting GBM’s reliability for individualized risk estimation. DCA demonstrated that GBM, RSF and XGBoost consistently yielded positive net benefit over both “treat none” and “treat all” reference strategies across a broad threshold range in the training set (Figures 4A, C and E). In the testing set, GBM sustained the greatest net benefit across the widest range of clinically plausible thresholds at all three timepoints (Figure 4B, D and F), with the advantage most pronounced at longer intervals where accurate SLOR stratification carries the greatest therapeutic relevance. Figure 4Net clinical benefit of nine machine learning models assessed at 12, 24 and 36 months in the training and testing sets of CD patients at risk of SLOR to IFX. (A) DCA results for the training set at 12 months. (B) DCA results for the testing set at 12 months. (C) DCA results for the training set at 24 months. (D) DCA results for the testing set at 24 months. (E) DCA results for the training set at 36 months. (F) DCA results for the testing set at 36 months.Graphs A-F compare net benefit against threshold probability (0-100%) for models: RSF, GBM, Lasso Cox, CoxBoost, SurvivalSVM, Deepsurv, Xgboost, SuperPC and plsRcox, plus ′Treat None′ and ′Treat All′. X-axis: ′Threshold Probability′; Y-axis: ′Net Benefit′. Models differ by line styles. Graph A: GBM and RSF excel at low thresholds; Xgboost and Deepsurv drop early. Graph B: GBM retains positive net benefit longest. Graph C: GBM and RSF dominate 10-50% thresholds. Graph D: GBM leads. Graph E: GBM and RSF show consistent benefits; others decline sharply. Graph F: GBM maintains highest net benefit across thresholds. ′Treat All′ crosses zero at 20-60% thresholds. GBM often provides superior net benefit across clinically relevant thresholds.Multi-line graphs (A-F) showing net benefit vs threshold probability for decision curve analysis models.Abbreviations: DCA, Decision curve analysis; CD, Crohn’s disease; SLOR, secondary loss of response; IFX, infliximab; RSF, random survival forest; GBM, gradient boosting machine; SVM, support vector machine; XGBoost, extreme gradient boosting; SuperPC, supervised principal component; plsRcox, partial least squares regression Cox model; Lasso-Cox, least absolute shrinkage and selection operator Cox regression; CoxBoost, Cox proportional hazards boosting; Deepsurv, deep survival neural network. Net clinical benefit of nine machine learning models assessed at 12, 24 and 36 months in the training and testing sets of CD patients at risk of SLOR to IFX. (A) DCA results for the training set at 12 months. (B) DCA results for the testing set at 12 months. (C) DCA results for the training set at 24 months. (D) DCA results for the testing set at 24 months. (E) DCA results for the training set at 36 months. (F) DCA results for the testing set at 36 months. DeLong pairwise testing formally contextualized these findings (Supplementary Figure 4). In the training set, RSF demonstrated statistically superior AUC relative to the majority of competing models across all three timepoints (P < 0.05); at 12 months, however, several models including Lasso-Cox, CoxBoost, GBM and plsRcox showed no statistically significant differences in pairwise comparisons among themselves, forming a statistically comparable subgroup at that interval. In the testing set, at 12 months, AUC differences among GBM, RSF, Lasso-Cox, CoxBoost, XGBoost and plsRcox were largely nonsignificant across most pairwise comparisons. At 24 and 36 months, GBM demonstrated statistically significantly higher AUC than RSF and XGBoost (P < 0.05), while differences between GBM and Lasso-Cox or CoxBoost remained nonsignificant, indicating that GBM, Lasso-Cox and CoxBoost constituted a top-performing tier. By contrast, SurvivalSVM and SuperPC demonstrated significantly inferior performance relative to the majority of competing algorithms at both 24 and 36 months (P < 0.05), clearly distinguishing them as lower-performing models. Against this background, GBM’s statistically favorable discrimination at longer timepoints, combined with comparatively reasonable calibration and broader clinical net benefit, collectively identified it as the most comprehensively reliable algorithm for predicting SLOR to IFX. Multi-Dimensional Analysis of Performance Metrics To provide a comprehensive evaluation beyond discrimination and clinical utility, model performance was further profiled across accuracy, sensitivity, specificity and F1 score at 12, 24 and 36 months (Supplementary Tables 4–6). Before detailing these metrics, the statistical power across time horizons warrants brief clarification. As established earlier, the overall multivariable Cox model was adequately powered at the cohort level, supporting stable estimation at 24 and 36 months. However, the 12 month analysis was inherently underpowered due to the limited SLOR events accrued within this early window, making performance estimates at this horizon comparatively more variable and directional. In the training set, RSF and XGBoost led across most metrics. RSF achieved AUCs of 0.884, 0.947 and 0.967, accuracy of 0.787, 0.872 and 0.892, and F1 scores of 0.508, 0.844 and 0.894 at 12, 24, and 36 months, respectively. XGBoost followed closely (AUCs 0.841, 0.896 and 0.944; accuracy 0.868, 0.807 and 0.870; F1 scores 0.569, 0.778 and 0.872), while GBM occupied an intermediate position (AUCs 0.790, 0.850 and 0.909; accuracy 0.812, 0.765 and 0.834; F1 scores 0.468, 0.730 and 0.844). The parsimonious Cox-based algorithms (Lasso-Cox, CoxBoost and plsRcox) produced closely overlapping profiles, with training-set AUCs ranging from 0.766 to 0.876 and F1 scores from 0.378 to 0.817. Upon internal validation in the testing set, the performance hierarchy underwent significant reorganization. RSF exhibited the most pronounced attenuation from training to testing (eg, testing AUCs dropping to 0.680, 0.798 and 0.831), indicative of its susceptibility to overfitting. Conversely, GBM emerged as the optimal model, achieving the highest testing-set AUCs at every horizon (0.696, 0.825 and 0.854 at 12, 24 and 36 months, respectively) along with a stable C-index of 0.767. We explicitly acknowledge that the simpler, highly portable Cox-based models delivered testing-set performance that was broadly comparable to that of GBM. This was particularly evident at the 12-month mark, where the limited event count constrained the statistical resolution among algorithms, resulting in tightly clustered testing AUCs (Lasso-Cox: 0.673; CoxBoost: 0.676; plsRcox: 0.676) that were only marginally lower than that of GBM. Although these simplified Cox proportional hazards-based alternatives achieved testing performance broadly comparable to GBM, several observations support GBM as the preferred prognostic model. First, GBM was the only algorithm to consistently rank first in testing-set AUC across all three prespecified time points, indicating stable, horizon-independent discrimination. Second, at the adequately powered 24- and 36-month horizons, GBM showed a clear advantage in both overall correctness and false-positive containment. Cox-based models tended to gain sensitivity at the expense of specificity (eg, Lasso-Cox yielded a sensitivity of 0.853 but a specificity of only 0.634 at 24 months), whereas GBM maintained robust specificity (0.846 at 24 months and 0.890 at 36 months) without materially compromising true-event detection (sensitivity 0.691 and 0.692, respectively). By maximizing overall accuracy and optimizing this balance, GBM effectively limits false-positive alarms while reliably identifying patients at genuine risk. Collectively, these metrics confirm GBM as the most balanced and robust algorithm for predicting long-term SLOR to IFX, while Cox-based models remain valuable alternatives when computational simplicity is prioritized. Risk Stratification and Survival Analysis Based on the Optimal GBM Model To evaluate the prognostic discrimination capacity of the finalized GBM model, all participants were dichotomized into low- and high-risk strata based on the median of model-predicted risk scores, with SLOR-free survival estimated via Kaplan-Meier analysis across a 36-month follow-up horizon. In the training set (Figure 5A), the two strata exhibited sharply divergent event-free trajectories from early timepoints. Low-risk patients (n = 223) sustained a SLOR-free probability exceeding 0.70 at month 36, whereas high-risk counterparts (n = 223) underwent a steep, near-continuous decline approaching 0.10 by the end of follow-up. Between-group separation was highly significant (HR = 7.757, 95% CI: 5.611–10.723; P < 0.001). Analogous findings were observed in the testing set (Figure 5B), where low-risk (n = 96) and high-risk (n = 95) patients demonstrated comparably divergent survival patterns. The corresponding HR of 6.462 (95% CI: 3.870–10.788; P < 0.001) confirmed robust risk stratification, with the moderately wider confidence interval reflecting the reduced testing sample size. Figure 5Kaplan-Meier curves depicting SLOR-free probability across model-predicted risk groups generated by the optimal GBM model in CD patients receiving IFX therapy. (A) Training set; (B) Testing set.Image A: Kaplan Meier graph shows event-free probability by risk group over 36 months. Low Risk curve remains near 1.00 until 6 months, then drops to 0.78 by 36 months. High Risk curve falls from near 1.00 after 6 months to 0.10 at 36 months. HR = 7.757 (5.611 to 10.723), P < 0.001. Number at risk: Low Risk 223 to 121; High Risk 223 to 21. Image B: Similar graph with Low Risk curve declining slightly to 0.80 at 36 months and High Risk curve dropping to 0.18. HR = 6.462 (3.870 to 10.788), P < 0.001. Number at risk: Low Risk 96 to 49; High Risk 95 to 13.Two Kaplan Meier plots of event free probability comparing low risk versus high risk groups over time.Abbreviations: IFX, infliximab; CD, Crohn’s disease; SLOR, secondary loss of response; HR, hazard ratio; GBM, gradient boosting machine. Kaplan-Meier curves depicting SLOR-free probability across model-predicted risk groups generated by the optimal GBM model in CD patients receiving IFX therapy. (A) Training set; (B) Testing set. Model Interpretability Through SurvSHAP Analysis To clarify the decision-making mechanism of the optimal GBM model, SurvSHAP analysis was applied to characterize cohort-wide predictor contributions from both aggregate and time-resolved perspectives. As depicted in Figure 6A, a pronounced hierarchy emerged among the seven predictors when ranked by mean absolute aggregated SurvSHAP values. IFX trough concentration during maintenance ranked first by a wide margin, with its overall contribution far exceeding those of all remaining variables. Disease duration ranked second, followed by ATI level during induction, concomitant IMM use, PLT, sex and WBC in descending order. Figure 6Model interpretability via SurvSHAP analysis in CD patients receiving IFX therapy. (A) Mean absolute aggregated SurvSHAP values ranking each predictor by its overall contribution to model output; (B) Longitudinal trajectories of mean absolute SurvSHAP values characterizing the time-varying influence of each feature throughout the observation period; (C) Beeswarm plot illustrating the cohort-wide distribution and directionality of patient-level aggregated SurvSHAP values across all predictors, with the color scale encoding normalized variable magnitude.Image A displays a horizontal bar chart with ′Variable′ on the Y-axis and ′Average aggregated SurvSHAP(t) value′ on the X-axis. Key variables include IFXtroughconcentrationduringmaintenance (0.14), Diseaseduration (0.02) and others with values around 0.01. Image B features a multi-line graph with ′time′ on the X-axis and ′Average SurvSHAP(t) value′ on the Y-axis. IFXtroughconcentrationduringmaintenance rises significantly to 0.24 by time 35, while other variables show minor increases. Image C presents a beeswarm scatter plot with ′Aggregated SurvSHAP(t) value′ on the X-axis and ′Variable′ on the Y-axis. IFXtroughconcentrationduringmaintenance points range from -0.25 to 0.22, while other variables cluster near 0.0 with varying spreads. A side scale labeled ′Variable value′ ranges from 0.00 to 0.75.A mixed chart showing SurvSHAP predictor ranking, time trajectories and a beeswarm distribution plot.Abbreviations: SurvSHAP, time-dependent SHapley Additive exPlanations for survival analysis; CD, Crohn’s disease; IFX, infliximab; ATI, antibodies to infliximab; IMM, immunosuppressant; PLT, platelet count; WBC, white blood cell count. Model interpretability via SurvSHAP analysis in CD patients receiving IFX therapy. (A) Mean absolute aggregated SurvSHAP values ranking each predictor by its overall contribution to model output; (B) Longitudinal trajectories of mean absolute SurvSHAP values characterizing the time-varying influence of each feature throughout the observation period; (C) Beeswarm plot illustrating the cohort-wide distribution and directionality of patient-level aggregated SurvSHAP values across all predictors, with the color scale encoding normalized variable magnitude. Temporal trajectory analysis (Figure 6B) further demonstrated that this dominance intensified progressively over the observation period. The SurvSHAP curve for IFX trough concentration during maintenance rose steeply and continuously, achieving unambiguous separation from all other features well before the midpoint of follow-up. Disease duration followed a comparable upward course at a substantially lower magnitude, whereas the remaining variables traced persistently flat trajectories, indicative of temporally stable and relatively minor contributions. Cohort-wide directional patterns were further clarified by the beeswarm plot (Figure 6C). Patients maintaining adequate IFX trough concentrations clustered toward positive aggregated values, reflecting a protective effect on sustained treatment response, while subtherapeutic levels correspondingly shifted distributions into negative territory. Higher induction-phase ATI burdens were similarly mapped to adverse contributions, consonant with established immunogenicity mechanisms. Disease duration and the remaining predictors (Concomitant IMM use, PLT, Sex, WBC) spanned considerably narrower distributional ranges, corroborating their ancillary modulatory roles. Collectively, these findings establish IFX trough concentration during maintenance as the dominant and temporally escalating determinant of SLOR risk, with disease duration, immunogenic burden and hematological indices providing supplementary yet clinically informative prognostic signals. Discussion This investigation presents the development and internal validation of a GBM survival model coupled with SurvSHAP to predict SLOR to IFX among patients with CD. In the testing set, the model demonstrated good discriminatory performance, with time-dependent AUC values of 0.696, 0.825 and 0.854 at 12, 24 and 36 months, respectively, along with reasonably calibrated probability estimates and favorable net benefit across DCA. Attribution analysis via SurvSHAP designated maintenance-phase IFX trough concentration as the predominant predictive contributor, with disease duration, induction-phase ATI, concomitant IMM use, PLT, sex and WBC ranking thereafter in decreasing order of importance. These observations collectively establish an interpretable, multimodal ML framework that unifies pharmacokinetic, clinical and laboratory information into a clinically applicable tool for patient-level SLOR risk stratification and timely treatment adjustment in CD. The present findings align with contemporary evidence affirming the multifactorial pathogenesis of SLOR to anti-TNF-α therapy, while extending prior work through enhanced multimodal integration and temporal interpretability. The seven independent predictors identified encompass pharmacokinetic, immunogenic, disease-related and host-intrinsic dimensions, each contributing distinct mechanistic insights. Maintenance-phase IFX trough concentration emerged as the most influential protective predictor (HR = 0.761, 95% CI: 0.715–0.810, P < 0.05). Sufficient circulating drug exposure sustains continuous neutralization of both soluble and membrane-bound TNF-α, thereby attenuating downstream NF-κB activation and restraining mucosal leukocyte recruitment.21 When trough concentrations fall below the therapeutic threshold, the elevated TNF-α burden enhances target-mediated drug disposition (TMDD), accelerating IFX clearance and exacerbating subtherapeutic exposure in the setting of active mucosal inflammation.22,23 These findings accord with evidence from multiple prospective cohorts establishing the primacy of maintenance trough levels for durable remission,4 collectively reinforcing the critical importance of proactive TDM in preventing SLOR. Closely coupled with trough pharmacokinetics, anti-drug immunogenicity represented another central determinant of treatment durability. Elevated ATI during induction ranked third in predictor importance (HR = 1.024, 95% CI: 1.012–1.037, P < 0.05). Immunogenicity arises when the chimeric murine variable domain of IFX triggers T-cell-dependent B-cell activation, yielding high-affinity IgG antibodies that form immune complexes subsequently degraded via neonatal Fc receptor (FcRn)-mediated endocytosis, thereby shortening IFX half-life and neutralizing its TNF-α binding site.6,24 HLA-DQA1*05 carriage has been identified as a principal genetic determinant of persistent ATI production,25 and quantitative ATI levels at induction completion constitute a practical biomarker for targeting early immunogenicity mitigation consistent with current TDM guidelines.26 Beyond these drug-related determinants, disease-intrinsic phenotypic features likewise modulated the durability of response. Longer disease duration independently increased the risk of SLOR to IFX. RCS analysis demonstrated an approximately linear increase in HR beyond a reference of 2 years (P for nonlinear = 0.087, suggesting an approximately linear association). Mechanistically, this association reflects progressive fibrostenotic remodeling and the accumulation of irreversible structural bowel damage associated with prolonged CD.27 Although anti-TNF agents such as IFX effectively neutralize TNF-α-mediated inflammation, their efficacy against established fibrosis and stricturing lesions remains limited.28 Consequently, patients with longer disease duration exhibit reduced capacity for sustained mucosal healing and are predisposed to earlier SLOR. In parallel with structural disease burden, systemic inflammatory indices also carried prognostic weight. Elevated PLT independently conferred increased risk (HR = 1.003, 95% CI: 1.001–1.004, P < 0.05). In CD, reactive thrombocytosis typically reflects ongoing systemic and mucosal inflammation, since pro-inflammatory cytokines such as interleukin-6 drive megakaryopoiesis and platelet production during the acute-phase response.29,30 Beyond functioning as a passive marker, activated platelets actively propagate intestinal inflammation by releasing mediators such as platelet factor 4, β-thromboglobulin, and soluble CD40 ligand, which facilitate leukocyte recruitment and amplify mucosal cytokine signaling.31,32 Persistently elevated PLT may therefore indicate a more aggressive inflammatory phenotype that is less amenable to sustained TNF-α blockade, thereby increasing the risk of SLOR. Nevertheless, the per-unit HR of 1.003 implies only a modest cumulative effect (approximately 1.16 for a 50 × 109/L rise), and the RCS analysis showed a linear trend without significant nonlinearity (P for nonlinear = 0.195). In SurvSHAP, PLT exhibited a narrow and temporally stable contribution profile. PLT therefore functions as a complementary chronic inflammatory marker within the integrated risk model rather than an independently actionable clinical threshold. Consistent with this inflammatory pattern, elevated baseline WBC also emerged as an independent predictor of SLOR, serving as a surrogate marker of persistent neutrophilic inflammation and heightened disease activity in CD. Mechanistically, two converging pathways may underlie this association. First, activated neutrophils express Fc-gamma receptors that facilitate sequestration and clearance of IFX-immune complexes via receptor-mediated elimination;33 Second, heightened disease activity amplifies TNF-α production, thereby accelerating TMDD of IFX and resulting in subtherapeutic trough concentrations and diminished long-term efficacy.23 Further support for this link comes from a two-sample Mendelian randomization study, which provided causal genetic-epidemiological evidence for the association between circulating leukocytes and IBD risk,34 lending additional mechanistic coherence to our findings. With respect to host factors, male sex independently increased SLOR risk (HR = 1.491, 95% CI: 1.048–2.120, P<0.05), a finding consistent with well-established sex-specific differences in IFX pharmacokinetics and immunogenicity. In a nationwide IBD cohort, male patients receiving IFX had significantly higher ATI concentrations and lower serum trough levels than their female counterparts.35 Population pharmacokinetic analyses further corroborate these observations, identifying male sex as an independent predictor of accelerated IFX clearance and an expanded volume of distribution.36 Against this backdrop of risk-amplifying factors, therapeutic co-interventions offered a counterbalancing protective influence. Concomitant IMM use emerged as a protective predictor in the multivariate model. The mechanisms by which IMM attenuates anti-TNF-α immunogenicity remain incompletely understood, though methotrexate and thiopurines are thought to reduce T-cell proliferation and suppress memory B cell and CD4+ T cell activity, thereby limiting ATI formation.37,38 By preserving IFX half-life and maintaining therapeutic trough exposure, combination therapy may attenuate the immunogenic disadvantage inherent to IFX monotherapy. The magnitude of this protective effect closely aligns with findings from the PANTS cohort and prior pharmacokinetic studies,4,39 supporting the biological plausibility of this association. Collectively, these mechanistic correspondences affirm the biological validity of the GBM model. Its principal advance over prior work resides in the simultaneous integration of all seven predictors within a temporally resolved survival framework, enabling the capture of cross-predictor interactions that static or single-timepoint models inherently cannot detect. SurvSHAP further enriches this framework by revealing how the relative dominance of individual predictors shifts across the follow-up horizon. Three findings diverged from a priori expectations. First, despite superior training-set performance, the RSF model exhibited greater validation-cohort attenuation than the GBM model. The regularized sequential ensemble construction of the GBM model confers stronger resistance to overfitting at moderate real-world sample sizes, consistent with the bias-variance tradeoff in decision-tree-based learners.14,40 Second, although CRP and ESR are widely regarded as standard inflammatory markers in routine CD monitoring,41,42 neither reached statistical significance in multivariable Cox regression nor contributed meaningfully to the final GBM model. This counter-intuitive result suggests that once pharmacokinetic, immunogenic and disease-chronicity parameters are comprehensively integrated, conventional systemic inflammatory indices may offer limited incremental prognostic value for long-term SLOR prediction. Third, WBC emerged as an independent predictor in multivariable Cox regression (HR = 1.022, P < 0.05) yet failed to attain overall significance in RCS analysis (P for overall = 0.159). This discrepancy most likely reflects the difference between an adjusted effect estimated within the training set and an unadjusted marginal association characterized across the full cohort, whereby WBC’s prognostic signal becomes statistically detectable only in the presence of correlated covariates. Integrating SurvSHAP into the GBM architecture transforms a black-box ensemble model into an interpretable clinical decision-support tool.43 By decomposing predicted SLOR risk into additive, time-indexed feature contributions, the framework supports both cohort-level predictor ranking and patient-specific temporal profiling. The predominance of IFX trough concentration during maintenance suggests the potential value of proactive TDM protocols, whereby levels below the exploratory RCS-derived threshold of 2.96 mg/L may warrant consideration of expedited dose escalation rather than reactive adjustment following clinical deterioration. Nevertheless, this threshold should be regarded as hypothesis-generating and requires prospective multicenter external validation before clinical implementation. Prominent induction ATI contributions advocate for early immunogenicity screening and inform decisions regarding IMM intensification or biologic switching.44,45 Baseline PLT and disease duration further guide risk-stratified monitoring frequency. Kaplan-Meier analysis confirmed clear risk-group separation in the testing set (HR = 6.406, 95% CI: 3.840–10.685, P < 0.001), offering a clinically actionable threshold for regimen modification. Collectively, this interpretable framework bridges advanced predictive analytics with point-of-care decision-making to support individualized biologic therapy management. Several methodological limitations merit careful consideration. First, the single-center retrospective design constitutes an inherent structural constraint. Despite rigorously standardized inclusion criteria, the specialized case-mix and treatment protocols of a tertiary referral center may not adequately reflect patient profiles and clinical practices across community-based or multicenter settings, potentially introducing referral bias and limiting the external applicability of the findings. Although benchmarking nine algorithms within a single institution provides a valuable methodological reference, reliance on data from one hospital inherently restricts the clinical generalizability of our results; we therefore explicitly acknowledge that the apparent advantage of the selected GBM model should be interpreted with caution until an independent external cohort from at least one additional center is available for retrospective back-testing and, ultimately, prospective validation. Second, the non-availability of fecal calprotectin, mucosal endoscopic scoring and patient-reported outcome data restricted thorough inflammatory phenotyping at critical assessment time points, potentially masking clinically meaningful disease activity signals. Third, the absence of pharmacogenomic profiling and gut microbiome characterization means that important sources of biological variability influencing IFX pharmacokinetics and immunogenic potential remain unaccounted for. Fourth, although stratified random splitting was applied, internal validation alone cannot fully establish generalizability, and prospective multicenter external validation remains essential for clinical translation. In addition, the credibility of the reported performance should be tempered by several considerations, including the modest early-phase discrimination, the optimistic training-set hazard ratio derived by resubstitution (HR = 7.757), and the comparable 12-month test-set performance of several Cox-based learners, with overlapping AUC confidence intervals and non-significant DeLong comparisons. These findings weaken any claim that GBM is unequivocally superior or ready for immediate deployment. We therefore frame our conclusions with deliberate restraint, emphasizing that prospective multicenter external validation must precede any clinical implementation. Collectively, these constraints outline a concrete agenda for future research. Multicenter prospective studies are warranted to externally validate the robustness and generalizability of the proposed GBM model across heterogeneous populations and clinical settings, mitigate single-center referral bias, and establish its clinical utility for individualized IFX therapy. Building on this, validation in geographically and ethnically diverse cohorts, together with iterative model updating as real-world evidence accumulates and the progressive incorporation of dynamic TDM data alongside emerging non-invasive biomarkers, will provide the necessary foundation for translation into routine clinical practice. Conclusion The interpretable GBM-based survival framework provides acceptable and temporally interpretable prediction of SLOR in CD by incorporating pharmacokinetic, immunological, inflammatory and demographic determinants grounded in established clinical and mechanistic evidence. The model demonstrated reasonable discriminatory performance, well-calibrated probability estimates and favorable net clinical benefit across DCA. Nonetheless, prospective multicenter external validation, together with systematic integration of dynamic TDM data, remains essential before clinical implementation.