Información complementaria
Background: Following pancreatoduodenectomy for distal cholangiocarcinoma (dCCA), R0 resection (i.e., tumor clearance of 1 mm between tumor and resection margin) is the strongest prognostic factor for overall survival (OS). Future randomized trials on neoadjuvant therapy in patients with dCCA could target patients at high risk of microscopic residual disease (R1) but large nationwide studies on how to identify such patients are lacking. This nationwide study aimed to identify preoperative predictors and develop a simple risk score for R1 resection in dCCA.
Methods: Nationwide retrospective cohort study including patients undergoing pancreatoduodenectomy for dCCA in the Netherlands Cancer Registry (2009–2023) with the primary aim to determine preoperative predictors for R1 resection, using multivariable logistic regression analyses.
Study design
A nationwide multicenter retrospective cohort study was conducted, using data from all sixteen centers performing PD from the population-based Netherlands Cancer Registry (NCR). The NCR systematically records all newly diagnosed malignancies in the Netherlands, after notification by the Dutch Nationwide Pathology Databank (PALGA) and the Dutch National Hospital Care Registration (LBZ, hospital discharges and outpatient visits). All notifications are verified in the electronic patient files in hospitals by trained NCR registrars approximately 9-12 months after diagnosis. Survival data are updated annually through linkage with the Dutch personal records database (last update 1 February 2025).
All consecutive patients (>18 years) who underwent PD for postoperatively histologically proven dCCA according to International Classification of Diseases for Oncology edition 4 topography C24.2, including rare tumor entities beyond adenocarcinoma in the Netherlands were included (2009-2023). Patients were excluded if residual disease status (R0 or R1) was unavailable, or if a surgical procedure other than PD was performed.
Statistical analysis
Univariable and multivariable logistic regression analyses were performed to identify independent preoperative predictors for R1 resections and estimate both unadjusted and adjusted odds ratios (ORs) and 95% confidence intervals (95% CIs). Several variables were either dichotomized (in case of categorical variables) or transformed (in case of continuous variables) to prevent sparse data bias and improve the fit of the model. ASA score, cT-stage, and cN-stage were dichotomized into: ASA 1-2 vs ASA 3-4, cT1-2 vs cT3-4, and N0 vs N1-2. Age, tumor size, and CA19-9 were winsorized to reduce the leverage of extreme values: age values <40 years were winsorized to 40 years, tumor size values >50 mm were capped at 50 mm, and CA19-9 values of 9000 U/mL and higher were winsorized to 9000 U/mL. Several continuous laboratory variables with skewed distributions (i.e., CA 19-9, pre-drainage bilirubin, and CRP) were natural log-transformed (ln) to improve the fit of the modeled covariate-outcome relationship. Missing data were handled using multiple imputation (50 imputations, 20 iterations) using multivariate imputation by chained equations with predictive mean matching.
An interaction analysis was conducted to account for changes in the pathological assessment over time. Specifically, since 2015, a more refined definition of resectability has been implemented; since 2016, a systematic pathological assessment protocol has been introduced; and since 2009, three different TNM classifications have been used.
In accordance with the TRIPOD+AI recommendations for prediction models, the predictive value of this model was evaluated in terms of its discrimination, calibration, and clinical utility. In brief, discrimination refers to how well the model can discriminate between patients with and without R1 margins, and was quantified using the area under the receiver operating characteristic curve (AUC). The AUC ranges from 0.5 (no discriminative power) to 1.0 (perfect discrimination); however, labelling systems for AUC values between these extremes (e.g., defining an AUC of >0.9 as ‘excellent’) are generally arbitrary and inadequate proxies of clinical utility. As such, AUC value intervals are presented as estimates with their corresponding confidence intervals without attempts to label the AUC value as, e.g., ‘fair’, ‘good’, or ‘excellent’. Calibration indicates the agreement between the predicted and observed risk of an R1 resection, and was quantified using flexible calibration curves, and the calibration intercept and slope (ideal values, respectively, 0 and 1). Lastly, decision curve analysis was performed to assess whether the clinical utility of a ‘model-based treatment approach’ was either equal to or superior compared to a ‘treat all’ and ‘treat none’ approach. This analysis was performed to conform with the TRIPOD+AI guidelines; however, as patient eligibility for surgery is dependent on several factors besides the expected risk of an R1 resection, the decision curve analyses are presented in the Supplementary Materials of the manuscript.
Internal validation was performed using bootstrapping, with 2000 bootstrap resamples in each imputed dataset, resulting in a total of 100.000 bootstrap datasets. Model training in each bootstrap sample followed all modeling steps (including backward selection with P<0.05) to accurately capture all model uncertainty and ‘researcher degrees-of-freedom’. Bootstrap-corrected performance measures were estimated using the approach of Efron and Gong. Bootstrapping was used for internal validation performs rather than split-sample validation (e.g., a training set of 70% and a validation set of 30%) or (repeated) cross-validation, as bootstrapping outperforms these methods in the estimation of bias-/overoptimism-corrected performance measures.
All statistical analyses were conducted using RStudio, version 4.5.1, with therms and mice packages.
Results: Among 1257 patients undergoing pancreatoduodenectomy for dCCA, an R1 resection occurred in 433 (34%). Independent preoperative predictors of R1 resection were: ASA score >2 (adjusted odds ratio, 1.82 [95% CI, 1.24 to 2.67]; P=0.004), vascular involvement on preoperative imaging (2.53 [1.56 to 4.09]; P=0.0001), clinical T-stage >2 (1.95 [1.04 to 3.63]; P=0.041), and clinical N-stage 1–2 (1.75 [1.20 to 2.53]; P=0.003). Other preoperative factors (i.e., sex, age, biliary drainage, CA19-9, CEA, and tumor size) were only significantly predictive for an R1 resection in univariable analyses. The prediction model for R1 resection showed an AUC of 0.71 (95% CI, 0.68 to 0.74) with excellent calibration (overoptimism-corrected intercept and slope of, respectively, –0.03 and 0.94). The risk of R1 disease ranged from 20% with no factors present to 80% with all factors present.
Conclusion: Prior to pancreatoduodenectomy for dCCA, patients at high risk of an R1 resection can be identified based on high ASA score, clinical T-stage, vascular and lymph node involvement on preoperative imaging. These predictors and the prediction model (readily available via pancreascalculator.com) may help to identify patients for neoadjuvant therapy in order to increase R0 resection rates.