Comparative performance of machine learning algorithms for predicting mortality among early-onset colorectal cancer...

Early-onset colorectal cancer (EOCRC), defined as colorectal cancer diagnosed before age 50, is increasing in the United States. Accurate mortality prediction may help identify high-risk patients and inform survivorship planning, yet the performance of machine-learning algorithms in state-level EOCRC populations remains unclear.

We conducted a retrospective population-based prediction study using Tennessee Cancer Registry data for adults aged 20-49 diagnosed with colorectal cancer from 2011 to 2020. The analytic cohort included 3467 patients with available survival time and vital status data. The primary outcome was all-cause mortality; secondary binary outcomes were 1-, 3-, and 5-year mortality. Predictors included age group, race, sex, stage, bone, brain, liver, and lung metastases, and comorbidity. Cox regression was used for survival modeling. Its discrimination was compared with two machine-learning models built for time-to-event data, a random survival forest and an XGBoost Cox model. Both were tested on the full cohort using random-split and temporal validation. Logistic regression, LASSO logistic regression, decision tree, random forest, and XGBoost were also evaluated for binary mortality prediction at 1, 3, and 5 years, using AUC, precision-recall AUC, accuracy, sensitivity, specificity, F1-score, calibration, variable importance, and subgroup performance.

Overall, 1064 patients died during follow-up. The Cox model showed good discrimination (C-index=0.79). The random survival forest and XGBoost Cox model performed about as well as the Cox model itself, with a C-index of 0.80 for all three under random-split validation and 0.77 under temporal validation. Stage and metastatic disease were the strongest mortality predictors. For 1-year mortality, random forest had the highest AUC (0.816). For 3-year mortality, XGBoost had the highest AUC (0.843). For 5-year mortality, XGBoost and logistic regression performed similarly, with AUCs of approximately 0.850. Variable-importance analyses consistently identified stage and metastatic disease as dominant predictors. Model performance varied by race, sex, and stage.

Machine-learning approaches, both binary classifiers and time-to-event models, did not outperform traditional regression, underscoring the value of transparent, interpretable prediction methods.