ResDeepSurv: A Survival Model for Deep Neural Networks Based on Residual Blocks and Self-attention Mechanism

Area of Science:

• Medical research
• Biostatistics
• Machine learning

Background:

• Survival analysis is vital in medicine for understanding disease progression and treatment effectiveness.
• Conventional models like Cox regression often require extensive feature engineering or prior knowledge.
• Personalized medicine necessitates models that can capture complex covariate-outcome relationships.

Purpose of the Study:

• To propose a novel residual-based self-attention deep neural network for survival modeling, named ResDeepSurv.
• To develop a model that simulates survival time distributions and covariate correlations without strict distributional assumptions.
• To offer a powerful alternative to traditional survival analysis methods, enhancing personalized medicine.

Main Methods:

• Developed ResDeepSurv, a deep neural network integrating self-attention mechanisms with Cox regression principles.
• The model simulates survival time distributions and covariate-outcome correlations, accommodating both linear and nonlinear risk functions.
• Validated performance on multiple publicly available clinical datasets.

Main Results:

• ResDeepSurv demonstrated performance on par with or superior to existing survival analysis methods across various risk functions.
• The model effectively captures complex relationships between covariates and survival outcomes.
• Superior performance was validated through evaluations on diverse clinical datasets.

Conclusions:

• ResDeepSurv proves effective for survival analysis, offering a promising alternative to traditional approaches.
• The model's ability to handle complex data without extensive feature engineering supports personalized medicine.
• ResDeepSurv enhances clinical decision-making by providing accurate survival predictions.