Continuous and discrete time survival analysis: neural network approaches.

Area of Science:

• Machine Learning
• Biostatistics
• Computational Biology

Background:

• Survival analysis is crucial for modeling time-to-event data in various scientific fields.
• Traditional survival models may have limitations in capturing complex data patterns.
• Neural networks offer a flexible approach for advanced statistical modeling.

Purpose of the Study:

• To develop and compare two novel neural network models for survival analysis.
• To evaluate the performance of continuous-time and discrete-time Bayesian neural network models.
• To assess the discrimination and calibration capabilities of these models on real-world data.

Main Methods:

• Bayesian inference framework for model training.
• Development of two neural network architectures: continuous-time and discrete-time formulations.
• Application and evaluation on a real-world survival analysis dataset.

Main Results:

• Both continuous and discrete time models showed good discrimination capabilities, with C-indices ranging from 0.75 to 0.81.
• The models exhibited good calibration performance (p<0.05) for predictions up to 7 years.
• Comparable performance was observed between the continuous and discrete time models.

Conclusions:

• The proposed Bayesian neural network models are effective for survival analysis.
• Both continuous and discrete time formulations provide valuable tools for time-to-event data modeling.
• These models offer robust discrimination and calibration, suitable for complex survival data.