Survival Analysis with High-Dimensional Omics Data Using a Threshold Gradient Descent Regularization-Based Neural...

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• High-dimensional omics data analysis with censored survival outcomes is increasingly common.
• Existing methods, often based on the Cox model, may lack flexibility for complex relationships like nonlinearity.
• Neural networks (NNs) show promise but are less developed for censored survival data, especially with high-dimensional omics measurements requiring regularization.

Purpose of the Study:

• To propose and evaluate a novel neural network (NN) approach for censored survival analysis using high-dimensional omics data.
• To introduce and integrate the threshold gradient descent regularization (TGDR) technique into NN models for improved performance and feature selection.
• To provide a more flexible and effective alternative to existing penalized NN methods in this domain.

Main Methods:

• Development of a TGDR-based neural network (NN) architecture specifically designed for censored survival data.
• Implementation of TGDR for regularized estimation and selection of relevant omics features.
• Comparative analysis against unregularized and penalized NN approaches through simulations.
• Validation on two real-world cancer omics datasets.

Main Results:

• Simulations demonstrate satisfactory performance of the proposed TGDR-based NN model.
• The TGDR technique shows competitive performance compared to penalization methods.
• The new NN architecture offers unique advantages over existing penalized and unregularized models.
• Practical effectiveness is confirmed through successful application to cancer omics datasets.

Conclusions:

• The TGDR-based NN provides a practical and effective new approach for survival analysis with high-dimensional omics data.
• This method enhances flexibility and regularization capabilities within the neural network paradigm for survival data.
• The study offers a valuable alternative for researchers dealing with complex omics and survival data challenges.