Survival stacking with multiple data types using pseudo-observation-based-AUC loss.

Area of Science:

• Machine Learning
• Biostatistics
• Survival Analysis

Background:

• Machine learning models struggle with right-censored survival data.
• Existing adaptations include pseudo-observations and inverse probability of censoring weighting (IPCW).
• Ensemble methods can enhance predictions, but integrating diverse survival algorithms is underexplored.

Purpose of the Study:

• To propose a novel ensemble procedure for improved risk prediction in right-censored survival data.
• To optimally stack predictions from various survival and survival-adapted machine learning algorithms.
• To evaluate the proposed ensemble method against existing strategies.

Main Methods:

• Developed an ensemble procedure utilizing the area under the pseudo-observation-based time-dependent ROC curve.
• Optimally stacked predictions from diverse survival algorithms (e.g., survival random forest, IPCW bagging, pseudo-observations).
• Applied and evaluated the method on real-world survival data.

Main Results:

• The proposed ensemble method demonstrated improved risk prediction performance.
• Outperformed single survival-based methods like survival random forest.
• Showed superiority over strategies relying solely on pre-processing steps like IPCW bagging or pseudo-observations.

Conclusions:

• The novel ensemble procedure effectively integrates diverse survival prediction algorithms.
• This approach offers a significant advancement in building accurate risk prediction models for censored survival data.
• The method provides a robust framework for leveraging ensemble learning in survival analysis.