PROGNOSTIC VALUE OF DEMOGRAPHICS, HISTOPATHOLOGY, AND MISMATCH REPAIR STATUS IN COLORECTAL CANCER: A TIME-TO-EVENT MACHINE LEARNING APPROACH

Tóm tắt

Background: Colorectal cancer prognosis depends on multiple factors including demographics, histopathology, and mismatch repair (MMR) status. Traditional Cox regression models have limitations in handling high-dimensional clinical data. Objectives: This study applied time-to-event machine learning algorithms to investigate the prognostic values of demographics, histopathology, and MMR status in predicting overall survival of colorectal cancer patients. Materials and Methods: A total of 165 colorectal cancer patients from Can Tho Oncology Hospital (2019-2021) were recruited. Input features included age, sex, pTNM stage, histological type, tumor grade, lymphovascular invasion (LVI), tumor-infiltrating lymphocytes (TILs), perineural invasion (PNI), and MMR protein status. The dataset was divided into training (70%) and testing sets. Five time-to-event algorithms were trained with 1000 bootstraps and hyperparameter tuning, then validated on the testing set. SurvSHAP package was used for feature importance ranking. Results: Gradient Boost Survival outperformed other models with acceptable discrimination (C-index: 0.812, 95% CI: 0.784-0.840) and calibration (integrated Brier score: 0.057, 95% CI: 0.056-0.058). Age, lymphovascular invasion, and MMR status were identified as the three most important predictive features. Feature importance increased during the first 24 months and then stabilized. Conclusion: Time-to-event ensemble machine learning models effectively predict survival prognosis in colorectal cancer. MMR status, combined with demographic and histopathological features, represents an important predictor of overall survival.