A Flexible Ensemble Learning Method for Survival Extrapolation.

Area of Science:

• Biostatistics
• Health Technology Assessment
• Machine Learning

Background:

• Survival extrapolation is crucial for long-term survival estimation using short-term clinical trial data, particularly in health technology assessment (HTA).
• Traditional methods rely on selecting single parametric models, which are highly sensitive to model misspecification, leading to potential inaccuracies.

Purpose of the Study:

• To propose a novel, robust method for survival extrapolation that overcomes the limitations of single-model selection.
• To enhance the accuracy and reliability of long-term survival estimates derived from clinical trial data.

Main Methods:

• Introduced Ensemble Learning for Survival Extrapolation (ELSE), an ensemble modeling approach.
• ELSE constructs a weighted average of estimates from multiple candidate models, utilizing nonparametric bootstrap for confidence intervals.

Main Results:

• Numerical simulations demonstrated ELSE's superior performance over traditional model selection methods like Akaike Information Criterion (AIC).
• Application to Therapeutically Applicable Research to Generate Effective Treatment Wilms Tumor (TARGET-WT) data showed improved point estimate accuracy and confidence interval coverage.

Conclusions:

• ELSE provides a robust and reliable method for survival extrapolation, performing well across various data models.
• The developed ensemble learning method offers enhanced real-world performance for survival estimation in HTA and clinical research.