Two-stage sampling for better survival model performance.

Area of Science:

• Biostatistics
• Bioinformatics
• Genomics

Background:

• High-dimensional censored survival data is increasingly used in health and medicine for risk prediction.
• Existing data splitting techniques for model training and evaluation lack examination regarding data splitting ratios and survival-specific characteristics in high-dimensional settings.

Purpose of the Study:

• To investigate the impact of data splitting ratios and survival-specific characteristics on survival model performance using high-dimensional censored data.
• To develop and validate an improved data sampling approach for enhanced survival model development.

Main Methods:

• Empirical study comparing simple random sampling and stratified sampling techniques on gene expression datasets using Lasso Cox models.
• Investigation of various data splitting ratios for simple random sampling and survival-specific variables for stratified sampling.
• Development and validation of a two-stage purposive sampling approach.

Main Results:

• Survival-specific characteristics significantly influence survival model performance across training, testing, and validation datasets.
• The proposed two-stage purposive sampling approach effectively mitigates excessive diversity in training data.
• This mitigation leads to improved survival model performance in both simulation and real-world data analyses.

Conclusions:

• Selecting appropriate sampling techniques and considering key factors are crucial for developing and validating survival models.
• The proposed two-stage purposive sampling method offers a viable solution for reducing data diversity and enhancing model performance.