An imputation approach using subdistribution weights for deep survival analysis with competing events

Area of Science:

• Computational Statistics
• Machine Learning in Healthcare
• Biostatistics

Background:

• Deep neural networks (DNNs) are increasingly used for survival data analysis, directly learning survival time distributions from predictors.
• Competing events, common in survival analysis, require specialized modeling due to their interdependencies and censoring.
• Existing DNN approaches for competing events often involve complex architectures with multiple subnetworks, posing training challenges.

Purpose of the Study:

• To propose a novel imputation strategy for deep neural networks (DNNs) to handle competing events in survival analysis.
• To simplify DNN architectures for competing events by avoiding the need for multiple event-specific subnetworks.

Main Methods:

• Developed a new data preprocessing imputation strategy incorporating weights from a time-discrete subdistribution hazard model.
• Applied this strategy to deep neural networks (DNNs) for survival data analysis with competing events.
• Compared the performance of the proposed method against traditional multi-subnetwork DNN architectures on synthetic and real-world datasets.

Main Results:

• The proposed imputation strategy effectively incorporates competing event information into DNNs.
• Experiments demonstrated that a single-event DNN design, using the new strategy, achieves comparable accuracy to multi-subnetwork models.
• The novel approach eliminates the need for complex, parameter-heavy DNN architectures for competing events.

Conclusions:

• The proposed imputation strategy offers a more efficient and simpler alternative for modeling competing events in DNN-based survival analysis.
• This method enhances the practicality and trainability of deep learning models for complex time-to-event data.
• The findings suggest a paradigm shift towards simplified DNNs for competing risks analysis.