Deep survival analysis for competing risk modeling with functional covariates and missing data imputation

Area of Science:

• Biostatistics
• Machine Learning
• Medical Informatics

Background:

• Accurate prognostic modeling is crucial in critical care for predicting patient outcomes.
• Traditional survival analysis methods struggle with complex, high-dimensional data and missing values.
• Competing risks, where multiple events can occur, complicate standard survival analysis.

Purpose of the Study:

• To introduce a unified deep-learning framework, the Functional Competing Risk Net (FCRN), for discrete-time survival analysis under competing risks.
• To develop a model that seamlessly integrates functional covariates and handles missing data within an end-to-end approach.
• To improve prediction accuracy in prognostic modeling, particularly in critical care settings.

Main Methods:

• Developed the Functional Competing Risk Net (FCRN), a deep learning framework for competing risks survival analysis.
• Integrated an Adaptive Learnable Basis (ALB) micro-network for functional data representation.
• Incorporated a gradient-based imputation module for simultaneous missing data handling and hazard prediction.

Main Results:

• FCRN demonstrated substantial improvements in prediction accuracy compared to random survival forests and traditional competing risks models.
• The model effectively handles functional covariates and missing data in a unified framework.
• Evaluated on simulated datasets and real-world ICU data (MIMIC-IV, Cleveland Clinic).

Conclusions:

• The Functional Competing Risk Net (FCRN) offers an advanced approach to prognostic modeling in critical care.
• FCRN effectively captures dynamic and static risk factors while accommodating irregular and incomplete patient data.
• This framework advances survival analysis by providing a robust method for complex, real-world healthcare data.