SurvBenchmark: comprehensive benchmarking study of survival analysis methods using both omics data and clinical data.

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Survival analysis is crucial for time-to-event data, but model performance evaluation is often limited.
• Existing comparisons primarily use clinical data and focus on prediction accuracy, neglecting censored omics data.
• Current methods for omics data often involve time binarization and classification, potentially losing information.

Purpose of the Study:

• To develop and present SurvBenchmark, a systematic framework for evaluating survival models.
• To compare classical and machine learning survival models using both clinical and omics data.
• To assess models based on multiple performance metrics beyond just prediction accuracy.

Main Methods:

• Developed SurvBenchmark, a systematic comparison design.
• Evaluated 20 survival models, including Cox and machine learning approaches, across 16 diverse datasets (clinical and omics).
• Assessed model performance using predictability, stability, flexibility, and computational efficiency.

Main Results:

• Model performance significantly varies across different datasets and chosen evaluation metrics.
• No single model consistently outperformed others across all scenarios.
• Highlighting the critical importance of employing multiple performance metrics for a balanced assessment.

Conclusions:

• A multi-metric evaluation is essential for a comprehensive understanding of survival model performance.
• Findings offer practical guidance for selecting appropriate survival analysis methods in translational research.
• Identified key areas for future research in survival modeling techniques and benchmarking strategies.