Two-stage sampling for better survival model performance

With the emergence of high-dimensional censored survival data in health and medicine, the use of survival models for risk prediction is increasing. To date, practical techniques exist for splitting data for model training and performance evaluation. While different sampling methods have been compared for their performances, the effect of data splitting ratio and survival specific characteristics have not yet been examined for high dimensional censored survival data.

We first conduct an empirical study of using the simple random sampling technique and stratified sampling technique on real high-dimensional gene expression datasets Lasso Cox model performance. For the simple random sampling technique, various data splitting ratios are investigated. For the stratified sampling, different survival specific variables are investigated. We consider C-index and Brier Score as evaluation metrics. We further develop and validate a two-stage purposive sampling approach motivated by our empirical study findings.

Our findings reveal that survival specific characteristics contribute to model performance across training, testing and validation data. The proposed two-stage purposive sampling approach performs well in mitigating excessive diversity within the training data for both simulation study and real data analysis, leading to better survival model performances.

We recommend careful consideration of key factors in different sampling techniques when developing and validating survival models. Using methods such as the proposed method to mitigate excessive diversity provides a solution.