Bayesian Networks Framework for Estimating Individual Survival Distributions: An Application in Amyotrophic Lateral Sclerosis Disease

Hamid Kalantari, Pouria Ramazi, Russell Greiner, Sanjay Kalra

Keywords: Survival analysis, Bayesian networks, Individual survival distribution, Amyotrophic Lateral Sclerosis

Abstract:

Accurate survival prediction is essential in many fields, particularly in healthcare, where estimating not only time-to-event but the entire individual survival distribution (ISD) can inform personalized decision-making. While existing models such as Cox proportional hazards, MTLR, and deep neural networks provide partial solutions, they often rely on strong parametric assumptions or lack interpretability. In this work, we propose a Bayesian network (BN)-based framework for survival prediction that explicitly models the joint distribution over covariates, survival time, and censoring status. We learn the structure of the BN using Gibbs sampling and parameterize the conditional distributions using flexible parametric models tailored to variable types.
Producing the ISD for an instance is performed by inference, which is done via sampling over the Markov blanket of survival and censoring nodes, enabling efficient estimation of personalized survival curves. We evaluate our method on a real-world (Amyotrophic Lateral Sclerosis) dataset and compare against state-of-the-art baselines using metrics such as C-index, MAE, and D-calibration.
Results show that our BN-based model offers competitive or superior performance in estimating full ISDs, while also providing interpretability.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 11451