Fast Lasso-type safe screening for Fine-Gray competing risks model with ultrahigh dimensional covariates

Area of Science:

• Biostatistics
• Survival Analysis
• Machine Learning

Background:

• The Fine-Gray proportional sub-distribution hazards (PSH) model is widely used for analyzing competing risks time-to-event data.
• Penalized regression methods, such as Lasso, are often employed to handle high-dimensional predictors in PSH models.
• Efficient feature selection is crucial for model interpretability and computational performance in such settings.

Purpose of the Study:

• To develop a fast and safe feature elimination method for penalized Fine-Gray PSH models.
• To ensure that the proposed method guarantees eliminated features are inactive in the final model.
• To evaluate the computational efficiency, screening performance, and prediction accuracy of the new method.

Main Methods:

• Development of the PSH-SAFE procedure for feature elimination.
• Integration of PSH-SAFE with Lasso and adaptive Lasso penalties for the Fine-Gray PSH model.
• Evaluation using simulated datasets and a real-world bladder cancer dataset.

Main Results:

• PSH-SAFE demonstrates desirable screening efficiency and safety properties.
• The procedure is computationally efficient and scales well to ultrahigh dimensional data.
• Empirical results show improved computational efficiency and comparable or better prediction performance than existing methods.

Conclusions:

• PSH-SAFE offers a practical and effective solution for feature selection in penalized Fine-Gray PSH models.
• The method enhances computational efficiency without compromising prediction accuracy or feature selection guarantees.
• PSH-SAFE is suitable for analyzing complex, high-dimensional time-to-event data with competing risks.