tdCoxSNN: Time-dependent Cox survival neural network for continuous-time dynamic prediction

Area of Science:

• Medical Imaging
• Biostatistics
• Machine Learning

Background:

• Dynamic prediction provides evolving, individualized risk assessments.
• Age-related macular degeneration (AMD) is a progressive eye disorder requiring accurate prognostic tools.
• Existing prediction models may not fully leverage longitudinal data, especially imaging.

Purpose of the Study:

• To propose a novel deep learning model, the time-dependent Cox survival neural network (tdCoxSNN), for dynamic prediction of AMD progression.
• To integrate longitudinal fundus images directly into a survival prediction framework.
• To evaluate the performance of tdCoxSNN against established methods.

Main Methods:

• Developed a time-dependent Cox survival neural network (tdCoxSNN) combining Cox survival models with deep learning.
• Utilized a convolutional neural network (CNN) within the tdCoxSNN to process longitudinal fundus images.
• Incorporated time-dependent covariates to capture evolving risk factors.
• Compared tdCoxSNN with joint modeling and landmarking approaches via simulations.

Main Results:

• The tdCoxSNN demonstrated commendable predictive performance in extensive simulation studies.
• The model achieved strong results when applied to the Age-Related Eye Disease Study (AREDS) dataset.
• The approach was also validated on a primary biliary cirrhosis dataset for predicting time-to-liver transplant.

Conclusions:

• The proposed tdCoxSNN effectively predicts disease progression using longitudinal imaging data.
• This deep learning approach enhances dynamic risk prediction for progressive diseases like AMD.
• tdCoxSNN offers a powerful tool for personalized, time-evolving risk assessment in clinical practice.