XGBoost Machine Learning Model Accurately Predicts Five-Year Colon Cancer Conditional Survival

An XGBoost machine learning model demonstrated superior performance in predicting five-year conditional survival for colon cancer patients compared to conventional logistic regression, according to a retrospective development and validation study. Conducted by researchers at Suzhou Municipal Hospital Baita Branch, the investigation established an optimized predictive framework leveraging both clinicopathological variables and dynamic longitudinal tumor biomarker measurements.

The study addressed a documented clinical gap, noting that most existing prognostic models for colon cancer are not tailored to the Chinese population and have not been implemented in routine clinical practice. To overcome these limitations, the investigators aimed to identify independent risk factors for five-year conditional survival and systematically compare conventional regression against several machine learning algorithms to establish an optimal predictive tool.

The retrospective study analyzed clinical records from 430 newly diagnosed colon cancer patients treated at Suzhou Municipal Hospital between 2015 and 2022. The dataset encompassed standard clinicopathological characteristics alongside dynamic post-diagnosis monitoring data for two primary tumor markers: carcinoembryonic antigen (CEA) and carbohydrate antigen 19-9 (CA19-9). To train and test the predictive architectures, patients were randomly divided in a 70:30 ratio into a training set of 301 individuals and a validation cohort of 129 individuals.

Feature selection was conducted using a dual algorithmic strategy combining the least absolute shrinkage and selection operator (LASSO) with the Boruta algorithm. This process identified nine core predictive factors from the candidate data: clinical stage, histological differentiation, distant metastasis, surgical treatment, targeted therapy, baseline CEA, baseline CA19-9, 12-month CEA, and 12-month CA19-9 levels.

Using these nine core predictors, the researchers constructed one conventional logistic regression model and six distinct machine learning architectures, including XGBoost and Random Forest algorithms. Model discrimination, calibration, and overall accuracy were evaluated across both cohorts using the area under the receiver operating characteristic curve (AUC), the Brier score, and the F1 score.

The XGBoost model achieved the best overall performance among all tested approaches, recording an AUC of 0.955, an F1 score of 0.904, and a Brier score of 0.085. Both the XGBoost and Random Forest models demonstrated predictive performance superior to that of the conventional logistic regression model, with XGBoost ultimately selected as the primary architecture due to its balance of discrimination, calibration, and integration with interpretability frameworks.

Model interpretability was evaluated using the SHapley Additive exPlanations (SHAP) framework to deliver quantitative, individualized risk interpretations. SHAP analysis identified 12-month post-treatment CEA level as the single most crucial predictor for survival prognosis, illustrating the clinical utility of dynamic biomarker monitoring in individualized disease assessment and ongoing management for patients with colon cancer.

The identification of nine core predictors highlighted a clear distinction between static clinicopathological factors and dynamic longitudinal biomarkers. In developing the predictive framework, the dual feature selection approach using the least absolute shrinkage and selection operator alongside the Boruta algorithm selected clinical stage, tumor differentiation, distant metastasis, surgical resection, and targeted therapy, together with baseline and twelve-month measurements of carcinoembryonic antigen and carbohydrate antigen 19-9. Rather than relying solely on fixed baseline prognostic characteristics, the incorporation of twelve-month post-treatment biomarker levels introduced a dynamic monitoring dimension to evaluate five-year conditional survival as patient risk profiles evolve following diagnosis and treatment.

The comparative modeling architecture systematically contrasted traditional statistical methodology against multiple machine learning approaches. The researchers developed one conventional logistic regression model and six separate machine learning models, including Random Forest and XGBoost algorithms, using the cohort of 430 newly diagnosed patients partitioned into 301 training cases and 129 validation cases. Across these implementations, both XGBoost and Random Forest demonstrated predictive performance superior to the conventional logistic regression model, illustrating that machine learning algorithms captured the predictive relationships among the nine clinicopathological and dynamic tumor marker variables more effectively than standard regression.

The selection of XGBoost as the primary model resulted from a comprehensive assessment across multiple performance dimensions rather than discrimination alone. The investigators evaluated all candidate models using the area under the receiver operating characteristic curve, the Brier score, and the F1 score. XGBoost achieved the leading performance profile among all tested approaches, recording an area under the curve of 0.955, the highest F1 score at 0.904, and the lowest Brier score at 0.085. This combination demonstrated an optimal balance between discriminative ability and probability calibration, while providing compatibility with the interpretability framework.

Model interpretability was achieved by integrating the SHapley Additive exPlanations framework, which enabled quantitative, individualized risk interpretations for each patient. The SHAP analysis revealed that twelve-month post-treatment carcinoembryonic antigen served as the crucial predictor for survival prognosis among all evaluated features. By providing individualized risk prediction methods and clarifying the role of risk factors, the model demonstrated the value of dynamic monitoring for disease assessment and disease management in colon cancer patients, addressing the absence of prognostic models tailored to the Chinese population.

Subject of Research: Machine learning prediction of five-year conditional survival in colon cancer

Article Title: A machine learning-based model for predicting 5-year conditional survival in patients with colon cancer: a development and validation study

Article References: Tu, Y., & Zeng, Y. (2026). A machine learning-based model for predicting 5-year conditional survival in patients with colon cancer: a development and validation study. Frontiers in Oncology, 16, Article 1962950. https://doi.org/10.3389/fonc.2026.1962950

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1962950

Keywords: colon cancer, machine learning, XGBoost, conditional survival, prognostic models, carcinoembryonic antigen, carbohydrate antigen 19-9, SHAP analysis, Random Forest, logistic regression, LASSO algorithm, Boruta algorithm, dynamic tumor markers, survival prognosis, clinical stage, distant metastasis, targeted therapy