Objective: Insidiousness is a hallmark of metachronous liver metastasis. Owing to the absence of a comprehensive machine-learning model integrating systemic inflammatory indicators for predicting metachronous liver metastasis after colon cancer surgery, and to the lack of prospective validation and interpretability, this study aimed to develop and validate a machine-learning model for predicting postoperative metachronous liver metastasis in patients with colon cancer. Methods: The variable pool of risk factors was determined through meta-analysis combined with three distinct screening approaches. The model was developed retrospectively and validated prospectively. The retrospective cohort comprised patients who underwent radical colectomy for colon cancer at X Hospital, S Hospital, and the P Hospital between 1 January 2012 and 1 January 2023. The prospective cohort included patients at X Hospital between 1 March 2023 and 1 August 2024. In the retrospective cohort, patients were randomly allocated to a training set and an internal validation set in a 7:3 ratio. Feature selection was performed using Lasso regression, multivariable logistic regression, and the Boruta random forest algorithm. The performance of ten machine-learning models was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score. AUC, calibration curves, and decision curve analysis were employed to elucidate clinical utility. Model interpretability was achieved through SHapley Additive exPlanations. This study strictly adhered to the TRIPOD + AI statement. Results: In retrospective cohort of 3938 patients, 11.2 % developed metachronous liver metastasis within 1 year; in prospective cohort of 724 patients, the corresponding proportion was 7.5 %. Following three feature-selection procedures and two multicollinearity assessments, 18 basic clinical variables and nine immune-inflammatory indices were selected for model development. Gradient boosting machine (GBM) demonstrated the highest overall performance, with an AUC of 0.964 (95 % CI: 0.944–0.983); compared with other models, decision-curve analysis revealed superior clinical utility. In the prospective cohort, the model maintained robust performance, achieving an AUC of 0.939. Conclusion: The GBM model demonstrated strong predictive performance and favorable clinical utility for identifying colon-cancer patients undergoing curative resection who are at risk of developing metachronous liver metastasis within 1 year. Multicenter model development followed by prospective validation underscored the clinical value of an integrated immunological signature. Early identification of high-risk patients permits intensified surveillance, timely intervention, and more efficient allocation of finite health care resources.
Kang et al. (Wed,) studied this question.