Survival prediction from imbalanced colorectal cancer dataset using hybrid sampling methods and tree-based...

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Colorectal cancer (CRC) has a high mortality rate (64.5%).
• Predicting CRC patient survival is vital for treatment decisions.
• Clinical data analysis for CRC survival prediction faces challenges, particularly with imbalanced outcomes.

Purpose of the Study:

• To develop algorithms for predicting 1-, 3-, and 5-year survival in colorectal cancer patients.
• To address the challenge of imbalanced datasets in survival prediction.
• To improve mortality prediction for the minority class of CRC patients.

Main Methods:

• Utilized a colorectal cancer dataset from the SEER database with imbalanced survival outcomes.
• Applied data pre-processing including handling missing values and merging low-frequency categories.
• Employed data balancing techniques: Edited Nearest Neighbor (ENN), Repeated Edited Nearest Neighbor (RENN), Synthetic Minority Over-sampling Technique (SMOTE), and SMOTE-RENN pipelines.
• Used tree-based classifiers: Decision Tree, Random Forest, Extra Tree, Gradient Boosting, and Light Gradient Boosting Machine (LGBM).
• Evaluated performance using 5-fold cross-validation.

Main Results:

• For 1-year survival prediction, the proposed method with LGBM achieved 72.30% sensitivity.
• For 3-year survival, RENN combined with LGBM yielded 80.81% sensitivity, demonstrating effectiveness on imbalanced data.
• For 5-year survival prediction, LGBM achieved 63.03% sensitivity.
• The RENN followed by SMOTE approach, with LGBM as the predictor, showed superior sensitivity for 1- and 3-year survival.
• LGBM outperformed other models in F1-score for the 5-year survival prediction task.

Conclusions:

• The developed algorithms effectively predict colorectal cancer survival, particularly in highly imbalanced scenarios.
• The proposed methods significantly enhance mortality prediction accuracy for underrepresented patient groups.
• The combination of data balancing techniques (RENN, SMOTE) and advanced classifiers (LGBM) offers a robust approach for CRC survival analysis.