Machine Learning-Based Model to Predict Survival in Resectable Intrahepatic Cholangiocarcinoma

Abstract

Purpose Many studies attempted to precisely predict survival or recurrence of patients with resectable intrahepatic cholangiocarcinoma using perioperative indicators. However, machine learning has been infrequently applied to predict postoperative survival in patients with intrahepatic cholangiocarcinoma (ICC). We therefore compared the performance of three prognostic models: the Cox proportional hazards model, random survival forest, and DeepSurv. Methods Least absolute shrinkage and selection operator (LASSO) regression was used for variable selection, and 13 variables were finally selected for model training. Model discrimination was evaluated using the concordance index (C-index) and time-dependent area under the receiver operating characteristic curve (time-dependent AUC). Model calibration was assessed using calibration curves and the integrated Brier score (IBS). Results A total of 394 consecutive patients with intrahepatic cholangiocarcinoma who underwent curative resection between January 2010 and December 2016 were retrospectively enrolled. In the validation set, the Harrell’s C-index was 0.680 for CoxPH, 0.700 for random survival forest, and 0.680 for DeepSurv. The random survival forest achieved the highest mean time-dependent AUC (0.738) and the lowest IBS (0.147), followed by DeepSurv (AUC: 0.723; IBS: 0.235) and CoxPH (AUC: 0.713; IBS: 0.148). The final model was visualized on a web-based platform (https://shizhengmi.shinyapps.io/exam/). Conclusion The random survival forest demonstrated favorable discrimination and calibration, highlighting the potential value of ensemble learning in intrahepatic cholangiocarcinoma. The top five predictors in the final model were CA19-9, albumin, tumor differentiation, CEA, and tumor number.