Prognostic Power of Ensemble Learning in Colorectal Cancer with Peritoneal Metastasis: A Multi-Institutional Analysis

Area of Science:

• Oncology
• Machine Learning
• Biostatistics

Background:

• Accurate survival prediction for colorectal cancer with peritoneal metastasis is challenging due to clinical heterogeneity.
• Traditional prognostic models have limitations in forecasting overall survival (OS).

Purpose of the Study:

• To evaluate machine learning boosting models against standard regression methods for OS prediction in colorectal cancer with peritoneal metastasis.
• To compare the performance of XGBoost and LightGBM against Ridge, Lasso, and linear regression.

Main Methods:

• Utilized a multi-institutional registry of 150 patients with synchronous peritoneal metastasis of colorectal cancer.
• Integrated 124 clinicopathological variables and employed preprocessing techniques like standardization and median imputation.
• Compared XGBoost and LightGBM against linear models using five-fold cross-validation and an XGBoost Cox model for right-censored data, with SHAP and LIME for interpretability.

Main Results:

• Boosting models significantly outperformed linear models, which exhibited high error rates and negative R2 values.
• XGBoost achieved a Mean Absolute Error (MAE) of 475 ± 60 and Root Mean Square Error (RMSE) of 585 ± 88.
• The XGBoost Cox model yielded a C-index of 0.64 ± 0.06, with SHAP analysis identifying inflammatory markers and peritoneal disease extent as key prognostic factors.

Conclusions:

• Boosting models offer a notable accuracy improvement over linear methods for survival prediction in this patient cohort.
• The prognostic power of current models remains moderate, indicating a need for further development.
• External validation is crucial before integrating these ensemble learning tools into clinical decision-making for colorectal cancer patients.