XGBLC: an improved survival prediction model based on XGBoost

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Survival analysis is vital for clinical research and disease therapy assessment.
• High-dimensional genomic data presents challenges for existing survival prediction models.
• Accurate prediction models are needed for genomic survival analysis.

Purpose of the Study:

• To propose an improved survival prediction model, XGBLC, for high-dimensional genomic data.
• To enhance survival prediction by integrating Lasso-Cox with the XGBoost framework.
• To improve the accuracy and reliability of survival analysis in cancer research.

Main Methods:

• Developed XGBLC, an XGBoost-based model incorporating Lasso-Cox.
• Defined novel first- and second-order gradient statistics for the XGBLC loss function.
• Validated XGBLC using simulated datasets and 20 cancer datasets from The Cancer Genome Atlas (TCGA) with 5-fold cross-validation.

Main Results:

• XGBLC demonstrated superior performance compared to five state-of-the-art methods on TCGA cancer datasets.
• The model achieved better C-index, Brier score, and AUC.
• XGBLC maintained accuracy and robustness on simulated datasets of varying scales.

Conclusions:

• XGBLC offers an accurate and robust approach for survival prediction using high-dimensional genomic data.
• The model can aid physicians in understanding genomic impacts on survival and personalizing cancer treatments.
• The XGBLC algorithm provides a valuable tool for genomic survival analysis.