Decision system for combined hepatocellular-cholangiocarcinoma distant metastasis: a retrospective cohort study based on machine learning

Combined hepatocellular-cholangiocarcinoma (CHC) is a rare primary liver cancer with low incidence and poor prognosis. CHC can develop distant metastasis (DM) at an early stage, which severely shortens the survival time of patients. We aimed to use machine learning (ML) to construct a decision-making system to assess risk factors for the development of DM in CHC patients. 1180 CHC patients from the SEER database between 2000 and 2020 were collected and randomized into the train set and internal test group in a 7:3 ratio. Patients diagnosed with CHC at our hospital between 2011 and 2018 were also collected as the external validation set ( N = 125). Patients with CHC were divided into a metastasis group and a non-metastasis group according to the occurrence of DM. Univariate and multivariate logistic regression analyses were conducted to evaluate the risk factors influencing DM of CHC patients in the training set. We screened the feature variables based on random forest (RF) and forward feature importance sequences. Then we incorporated them into six ML algorithms for constructing machine learning models and used 10-fold cross-validation for internal validation. Receiver operating characteristic (ROC) curves, precision-recall curve (PRC), calibration curve (CC), and confusion Matrix (CM) were used to evaluate the predictive ability of the models. We determined the importance ranking of risk factors for DM in CHC patients using Shapley additive explanations (SHAP) and further used SHAP to analyze the interpretability of the black-box model. Finally, we built a web risk calculator based on the optimal performance model to facilitate its clinical application. A total of six variables were selected for constructing the decision model. Extreme gradient boosting (XGB) obtained the most significant recognition ability, with ROCAUC = 0.863, accuracy of 0.802, sensitivity of 0.875, and PRAUC of 0.642 in the internal test set. 10-fold cross-validation results showed that the ROCAUC of XGB was 0.989, with a standard error of 0.019. The SHAP method revealed that node, surgery, age, grade, primary site, and race are the top 6 key variables that contribute to the occurrence of DM in CHC. In addition, the analysis of two typical cases proved the reliability of the model. The XGB model outperforms other machine learning methods in recognizing the occurrence of distant metastases in CHC patients, and has a high degree of utility and reliability to inform clinical treatment decisions for patients.