Improved nonparametric survival prediction using CoxPH, Random Survival Forest & DeepHit Neural Network

Area of Science:

• Bioinformatics
• Biostatistics
• Computational Biology

Background:

• High-dimensional bioinformatics data present challenges for classical survival models, leading to overfitting and low prediction accuracy.
• Traditional methods struggle with time-to-event data and complex covariate landscapes common in biological research.

Purpose of the Study:

• To propose and evaluate a novel hybrid feature selection approach for improved survival prediction in high-dimensional bioinformatics datasets.
• To enhance the reliability and robustness of variable selection for more accurate survival analysis.

Main Methods:

• Explored four variable selection techniques: LASSO, RSF-vs, SCAD, and CoxBoost for non-parametric biomedical survival prediction.
• Employed survival models (CoxPH, RSF, DeepHit NN) using selected variables.
• Introduced a novel method selecting variables consistently identified by a majority of the initial techniques.

Main Results:

• The proposed hybrid method demonstrated superior performance compared to individual feature selection techniques.
• Evaluated using Integrated Brier Score (IBS), Concordance Index (C-Index), and Integrated Absolute Error (IAE) on high-dimensional survival datasets.
• Real-world data applications confirmed the proposed method's enhanced survival prediction accuracy.

Conclusions:

• The proposed hybrid feature selection strategy offers a more robust and accurate approach for survival prediction with high-dimensional bioinformatics data.
• This method effectively addresses the limitations of classical models and individual feature selection techniques.
• The findings suggest significant potential for improving clinical outcome predictions in genomics and related fields.