Survival time prediction by integrating cox proportional hazards network and distribution function network

Area of Science:

• Biostatistics
• Machine Learning in Healthcare
• Survival Analysis

Background:

• The Cox proportional hazards model is standard for predicting hazard ratios but cannot directly generate individual survival times.
• Traditional survival analysis requires selecting specific distributions (e.g., exponential, Weibull) to estimate survival times from hazard ratios.

Purpose of the Study:

• To develop a novel method for predicting individual survival times by integrating hazard and distribution function networks.
• To overcome the limitations of the Cox model in directly generating survival time predictions.

Main Methods:

• Adapted the Cox proportional hazards model into a DeepSurv network for hazard ratio prediction.
• Developed a distribution function network to generate survival time predictions.
• Introduced a new evaluation metric based on intersection over union (IoU) for curve comparison.
• Utilized 1D gradient-weighted class activation mapping (Grad-CAM) for visualizing prognostic factor importance.

Main Results:

• The integrated network approach effectively predicts individual survival times.
• The proposed method demonstrated superior performance compared to existing methods.
• The IoU metric provided a robust evaluation of survival curve predictions.
• Grad-CAM visualizations highlighted key prognostic factors.

Conclusions:

• The combination of Cox proportional hazards network and distribution function network effectively generates accurate survival time predictions.
• This integrated approach offers a significant advancement in survival analysis and personalized medicine.