[Interpretable machine learning-based predictive model for assessing abdominal surgery risk and biologic therapy efficacy in CDAI 0 to 1 level Crohn disease patients].

OBJECTIVES

Crohn disease (CD) patients face a clinically significant high risk of abdominal surgery. This study aims to develop a predictive model for estimating abdominal surgery in CD patients with Crohn Disease Activity Index (CDAI) 0-1.

METHODS

CD patients treated at the Second Xiangya Hospital of Central South University between 2016 and 2022 were retrospectively enrolled. Using a fixed random seed, the full cohort was randomly split into a training set and a validation set at a 5:5 ratio. Final predictors were selected using multivariable backward stepwise Cox regression, and hazard ratios (HRs) with 95% confidence intervals (CIs) were estimated for surgery-associated factors. In addition, 8 machine learning survival prediction models were introduced, including least absolute shrinkage and selection operator (LASSO)-Cox regression, random survival forest (RSF), gradient boosting machine (GBM), CoxBoost, survival support vector machine (SurvivalSVM), extreme gradient boosting (XGBoost), supervised principal component regression (SuperPC), and partial least squares Cox regression (PLSR-Cox). The model with the best concordance index (C-index) performance in the validation set was chosen as the final decision-support model. Model discrimination was evaluated using the C-index and time-dependent ROC curves with corresponding area under the curve (AUCs), while calibration was assessed using calibration curves. Clinical net benefit was quantified using decision curve analysis (DCA). Shapley additive explanations (SHAP) was applied to interpret the contribution of model features to prediction, improving model explainability. Kaplan-Meier curves were used to describe cumulative surgery-free probability, and the Log-Rank test was used to compare differences across predicted risk strata and between biologic-exposed and non-exposed cohorts.

RESULTS

A total of 615 patients were included in the study, comprising 307 patients in the training set and 308 patients in the validation set. Multivariable backward stepwise Cox regression identified 4 key variables significantly associated with abdominal surgery risk in CD patients with CDAI 0-1, including C-reactive protein (CRP, HR=1.07, 95% CI 1.01 to 1.14, P=0.025), albumin (ALB, HR=0.69, 95% CI 0.46 to 1.04, P=0.075), fibrinogen (Fg, HR=0.65, 95% CI 0.51 to 0.84, P<0.001), and Montreal B behavior classification (HR=2.26, 95% CI 1.23 to 4.17, P=0.009). Among 9 candidate predictive models, CoxBoost achieved the best performance in the validation set, yielding a C-index of 0.746 (95% CI 0.683 to 0.809). Time-dependent ROC analysis demonstrated that in the training set, the 1-, 3-, and 5-year AUCs were 0.778, 0.749, and 0.772, respectively, while in the validation set, the corresponding AUCs were 0.761, 0.797, and 0.751. Calibration curves showed good agreement between predicted and observed surgery risk, with Brier scores <0.25 at all evaluated time points. DCA showed that CoxBoost provided clinical net benefit within a meaningful probability range. SHAP analysis result indicates that Montreal B subtype is the primary factor influencing the modeled predictive outcomes. Risk stratification generated by the CoxBoost model revealed that the cumulative surgery-free probability differed significantly among low-, moderate-, and high-risk groups (Log-Rank P<0.001). In the moderate-to-high-risk group, exposure to biologic therapy was associated with a significant reduction in abdominal surgery risk (HR=0.54, 95% CI 0.33 to 0.89, P=0.014). In contrast, no significant effect of biologic therapy was observed in the low-risk group (HR=1.01, 95% CI 0.51 to 2.00, P=0.970).

CONCLUSIONS

The CoxBoost model developed in this study effectively predicts abdominal surgery risk in CD patients with CDAI 0-1, and supports clinical decision-making regarding biologic therapy, providing evidence for personalized treatment planning.