Predictive model for very early recurrence of patients with perihilar cholangiocarcinoma: a machine learning approach
Jun Kawashima1,2, Yutaka Endo3, Zayed Rashid1
1Department of Surgery, The Ohio State University Wexner Medical Center and James Comprehensive Cancer Center, Columbus, OH, USA.
Background:
Although offering the best chance of potential cure for patients with localized perihilar cholangiocarcinoma (pCCA), resection has been associated with high morbidity and sometimes poor long-term outcomes due to recurrence. We sought to develop a predictive model to identify individuals at high risk for very early recurrence (VER) after curative-intent surgery for pCCA.
Methods:
Patients who underwent curative-intent surgery for pCCA between 2000-2023 were identified from a multi-institutional database. An eXtreme Gradient Boosting (XGBoost) model was developed to estimate the risk of VER, defined as recurrence within 6 months after resection. The relative importance of clinicopathologic factors was determined using SHapley Additive exPlanations (SHAP) values.
Results:
Among 434 patients undergoing curative-intent resection for pCCA, 65 (15.0%) patients developed VER. Median overall survival (OS) among patients with and without VER was 8.4 [interquartile range (IQR) 6.6-11.3] versus 38.5 (IQR 31.9-45.7) months (P<0.001). An XGBoost model was able to stratify patients relative to the risk of VER [low-risk: 6-month recurrence-free survival (RFS) 94.6% vs. intermediate-risk: 6-month RFS 88.3% vs. high-risk: 6-month RFS 40.0%; P<0.001]. Similarly, 3-year OS incrementally worsened based on VER risk (low-risk: 75.3% vs. intermediate-risk: 19.5% vs. high-risk: 4.6%; P<0.001). The SHAP algorithm identified age, preoperative carbohydrate antigen 19-9 (CA19-9) levels, tumor size and differentiation/grade, as well as lymph node metastasis as the five most important predictors of VER. The predictive accuracy of the model was good in the training [c-index: 0.74, 95% confidence interval (CI): 0.67-0.81] and internal validation (c-index: 0.77, 95% CI: 0.71-0.83) cohorts. An easy-to-use risk calculator for VER was developed and made available online at: https://junkawashima.shinyapps.io/VER_hilar/.
Conclusions:
A novel, machine learning based model was able to predict accurately the chance of VER after curative-intent resection of pCCA. In turn, the tool may help surgeons in the selection of patients likely to benefit the most from resection, as well as counsel individuals about the anticipated risk of recurrence in the early post-operative period.
