Deep Gated Neural Network With Self-Attention Mechanism for Survival Analysis

Area of Science:

• Biostatistics
• Machine Learning
• Medical Informatics

Background:

• Survival analysis models event occurrences, often relying on restrictive assumptions like proportional hazards.
• Existing methods may not fully capture temporal patterns in covariate effects, limiting predictive performance.

Purpose of the Study:

• To propose a novel framework, SA-DGNet, for survival analysis that overcomes limitations of traditional methods.
• To enhance risk prediction for single and competing risks by incorporating time-dependent and nonlinear covariate effects.

Main Methods:

• Developed a deep gated neural network with a self-attention mechanism (SA-DGNet).
• Transformed survival analysis into a time-series forecasting problem, treating time as an input covariate.
• Integrated multi-scale time-aware and scaled dot-product self-attention for improved data perception.

Main Results:

• SA-DGNet demonstrated superior performance compared to state-of-the-art methods on real-world datasets.
• The framework effectively models time-dependent and nonlinear covariate effects without distributional assumptions.
• Achieved significant improvements in survival analysis and risk prediction.

Conclusions:

• Gated neural networks and self-attention mechanisms show significant potential in survival analysis.
• SA-DGNet offers an effective approach for risk prediction using structured data, particularly in medical applications.
• The proposed method advances the field by removing traditional assumptions and capturing complex temporal dynamics.