deepAFT: A nonlinear accelerated failure time model with artificial neural network.

Area of Science:

• Computational Statistics
• Machine Learning in Survival Analysis

Background:

• Traditional survival analysis models like Cox regression assume linear relationships between covariates and survival time.
• These linear assumptions limit their ability to capture complex, nonlinear covariate effects in survival data.
• Accelerated failure time (AFT) models offer a flexible framework but often rely on parametric assumptions.

Purpose of the Study:

• To propose novel nonparametric, nonlinear algorithms (deepAFT methods) for survival outcome modeling within the AFT framework.
• To develop deep learning-based methods capable of directly predicting survival outcomes and handling censored data.
• To evaluate the prediction accuracy and robustness of deepAFT methods compared to existing survival models.

Main Methods:

• Developed deep artificial neural network (deepAFT) algorithms for survival outcome prediction.
• Employed imputation, re-weighting, and inverse probability of censoring weighting techniques to address data censoring.
• Validated methods through extensive simulation studies and application to a lymphoma clinical trial dataset.

Main Results:

• DeepAFT methods demonstrated accurate survival outcome predictions, outperforming traditional regression models.
• Achieved high prediction accuracy comparable to established deep learning survival methods (deepSurv, random survival forest).
• Successfully modeled nonlinear covariate effects and provided survival and cumulative hazard functions without extra learning.

Conclusions:

• DeepAFT methods provide a flexible and robust alternative to Cox regression for survival analysis, especially with nonlinear covariate effects.
• The proposed deep learning approach effectively handles censored data and offers superior prediction accuracy.
• DeepAFT methods are valuable tools for situations where standard survival models may be inadequate.