An empirical approach to model selection through validation for censored survival data

Area of Science:

• Biostatistics
• Medical Informatics
• Clinical Epidemiology

Background:

• Prognostic models are crucial for predicting disease progression, but existing variable selection methods struggle with bias in time-to-event data.
• Optimizing predictive performance often requires balancing model bias and variance, a challenge not fully addressed by current techniques.

Purpose of the Study:

• To propose a novel, two-stage variable selection approach combining Stepwise Tuning in the Maximum Concordance Index (STMC) and Forward Nested Subset Selection (FNSS).
• To enhance prognostic model performance by reducing estimation and selection bias in right-censored time-to-event data.

Main Methods:

• A two-stage variable selection process: Stage 1 uses STMC with cross-validation for optimism correction to identify optimal risk factor subsets.
• Stage 2 employs another selection method on intermediate results to mitigate overfitting and select a final parsimonious model.

Main Results:

• The proposed STMC-FNSS approach demonstrated superior performance in selecting improved and reduced average models compared to traditional methods (stepwise, AIC, lasso).
• Case studies and simulations showed the method yields better final models than full models across various performance measures.
• Independent validation confirmed the enhanced predictive performance of the selected models.

Conclusions:

• The novel two-stage variable selection method effectively improves prognostic model accuracy and generalizability for time-to-event data.
• This approach offers a robust strategy for selecting parsimonious and high-performing models in clinical prediction.
• The STMC-FNSS method provides a valuable tool for researchers aiming to optimize prognostic model development.