What question did this study set out to answer?

This research aims to create a machine learning model to predict metachronous liver metastasis in colorectal cancer patients to enhance early diagnosis and treatment.

July 29, 2026Scientific Reports0 citationsOpen Access

Interpretable machine learning models for predicting the risk of metachronous colorectal liver metastases

KWKejia Wu HXHonghu Xie CSChangbing Shen

Key Points

This research aims to create a machine learning model to predict metachronous liver metastasis in colorectal cancer patients to enhance early diagnosis and treatment.
Conducted a retrospective analysis of 620 colorectal cancer patients post-surgery with data split into training and validation cohorts.
Applied machine learning models such as Logistic Regression, Random Forest, LightGBM, XGBoost, and SVM with KNN imputation for missing values.
Developed a dynamic online nomogram for risk visualization based on logistic regression features.
The final logistic regression model showed the best performance with significant risk factors including T stage, confirming good ROC analysis results.
The model indicated a clinical benefit and impact within the relevant probability threshold for metachronous liver metastasis risk.
SHAP analysis highlighted T stage as the most influential feature in predicting liver metastasis.

Abstract

Abstract Liver metastasis is a frequent complication in colorectal cancer (CRC), significantly impacting patient prognosis. This study aims to develop a machine learning-based prediction model for metachronous liver metastasis (MLM) in CRC patients, facilitating early diagnosis and intervention to potentially improve treatment outcomes and survival rates. A retrospective analysis was conducted on 620 consecutive patients who underwent radical colorectal cancer resection at the First People’s Hospital of Changzhou during the study period and met the predefined inclusion and exclusion criteria. MLM status was determined according to postoperative follow-up outcomes rather than used as a sampling criterion. Among the final eligible primary cohort, 373 patients had no observed liver metastasis during follow-up, whereas 247 patients were diagnosed with metachronous liver metastasis more than 6 months after radical CRC resection. Patients were split into training (non-MLM = 258, MLM = 176) and internal validation (non-MLM = 258, MLM = 71) cohorts, with an external cohort of 52 non-MLM and 30 MLM patients from the Seventh People’s Hospital of Changzhou. Missing values were imputed using KNN. Based on the features selected by Logistic Regression (LR) and LASSO, five machine learning models (LR, RF, LightGBM, XGBoost, and SVM) were developed. Model performance was comprehensively evaluated using AUROC, DCA, accuracy, sensitivity, specificity, and F1 score, with SHAP illustrating feature influence. The best model was validated internally and externally. Eight independent risk factors (Age, Tumor Embolus, Size, T stage, N stage, Tumor Differentiation, RDW, AST) were included as features in the model. The LR model demonstrated the best performance, with SHAP indicating T stage as the most influential factor. A dynamic online nomogram was constructed to visualize the LR model. ROC analysis showed good discriminative performance of the final LR model, and calibration analysis suggested acceptable agreement between predicted and observed MLM risk. DCA and CIC analyses suggested potential clinical net benefit and clinical impact within the prespecified clinically relevant threshold probability range. In this study, LR model was selected as the final model for predicting the risk of metachronous liver metastasis after radical CRC resection. The dynamic nomogram provides an interpretable and accessible tool for visualizing individualized risk estimates. Given the retrospective design and current validation limitations, the model should be regarded as a supplementary tool to support postoperative risk stratification and follow-up planning rather than as definitive evidence for clinical decision-making. Further prospective and multicenter validation is warranted before routine clinical implementation.

Ask AI

Helpful

Bookmark

View Full Paper