Bayesian parametric models for survival prediction in medical applications

Area of Science:

• * Medical Statistics
• * Computational Biology
• * Bioinformatics

Background:

• * Personalized medicine requires patient-level treatment benefit prediction.
• * Survival models are crucial for time-to-event data, incorporating censoring.
• * Bayesian models offer advantages in handling uncertainty and local data adaptation.

Purpose of the Study:

• * Evaluate Bayesian parametric survival models against conventional methods.
• * Assess model performance on diverse public datasets (cardiology, infectious diseases, oncology).
• * Investigate model updating capabilities using Bayes rule for privacy-preserving adaptation.

Main Methods:

• * Implemented Bayesian parametric survival models (Exponential, Weibull) in Python.
• * Utilized linear combination and neural networks for parameter prediction.
• * Employed superiority and equivalence designs for comparative analysis against Cox PH, RSF, and NN-Cox PH.

Main Results:

• * Bayesian models demonstrated performance comparable to state-of-the-art methods.
• * Bayesian models required fewer hyperparameters and exhibited reduced overfitting.
• * Model updating via Bayes rule achieved performance equivalent to retraining on combined datasets.

Conclusions:

• * Bayesian parametric survival models are non-inferior to conventional survival models.
• * These models offer reduced hyperparameter tuning, less overfitting, and efficient updating.
• * Bayesian models provide crucial uncertainty quantification for statistical inference and prediction.