Survival analysis on rare events using group-regularized multi-response Cox regression.

Area of Science:

• Biostatistics
• Computational Biology
• Genomics

Background:

• Cox proportional hazard models struggle with prediction accuracy when training data has few uncensored events.
• High-dimensional survival data, common in large biobanks, often presents challenges due to sparse event occurrences.

Purpose of the Study:

• To develop a robust method for improving prediction performance in large-scale, high-dimensional survival data with few observed events.
• To leverage related survival responses with abundant events to enhance prediction for rare event outcomes.

Main Methods:

• A Sparse-Group regularized Cox regression approach was proposed.
• The method integrates multiple survival responses sharing common predictors, particularly useful for datasets like the UK Biobank.
• An accelerated proximal gradient optimization algorithm and a screening procedure were developed for scalability.

Main Results:

• The proposed method aims to achieve higher prediction performance compared to analyzing individual rare event responses separately.
• The approach is designed to be practical for large-scale datasets, addressing computational challenges.
• The developed optimization and screening procedures facilitate efficient analysis.

Conclusions:

• The Sparse-Group regularized Cox regression offers a promising solution for prediction tasks involving rare events in high-dimensional survival data.
• This method enhances the utility of large biobank datasets by enabling joint analysis of common and rare diseases.
• The computational advancements ensure the method's applicability to real-world, large-scale biological and medical research.