Deep Neural Network-Based Accelerated Failure Time Models Using Rank Loss.

Area of Science:

• Biostatistics and Survival Analysis
• Machine Learning in Healthcare
• Statistical Modeling

Background:

• Accelerated Failure Time (AFT) models offer intuitive covariate interpretation compared to hazard-based survival models.
• Semiparametric AFT models provide flexibility by not assuming specific error distributions, making them robust.
• Existing AFT models often assume linear predictors, limiting their ability to capture complex relationships in failure time data.

Purpose of the Study:

• To propose a novel deep neural network (DNN) approach for fitting Accelerated Failure Time (AFT) models.
• To investigate the performance of the proposed DNN-based AFT model, termed DeepR-AFT, particularly in scenarios with non-linear predictors.
• To compare the DeepR-AFT model against traditional parametric and semiparametric AFT models.

Main Methods:

• Application of deep neural networks (DNNs) to fit AFT models.
• Utilization of a Gehan-type loss function combined with a sub-sampling technique for model training.
• Extensive simulation studies to evaluate finite sample properties and comparative performance of the DeepR-AFT model.

Main Results:

• The DeepR-AFT model demonstrated superior performance compared to parametric and semiparametric AFT models when predictor relationships were non-linear.
• For linear predictors, DeepR-AFT showed improved performance, especially in high-dimensional covariate settings.
• Validation of the proposed model's effectiveness using three real-world datasets.

Conclusions:

• Deep neural networks provide a powerful tool for extending AFT models to handle non-linear predictor effects.
• The DeepR-AFT model represents a significant advancement for survival data analysis, offering improved accuracy and flexibility.
• The proposed method is effective for both non-linear and high-dimensional linear predictor scenarios in survival analysis.