Abstract
We are interested in survival analysis of hemodialysis patients for whom several biomarkers are recorded over time. Motivated by this challenging problem, we propose a general framework for multivariate joint longitudinal-survival modeling that can be used to examine the association between several longitudinally recorded covariates and a time-to-event endpoint. Our method allows for simultaneous modeling of longitudinal covariates by taking their correlation into account. This leads to a more efficient method for modeling their trajectories over time, and hence, it can better capture their relationship to the survival outcomes.
Introduction
In clinical research, particularly concerning chronic conditions such as end-stage renal disease, patients are monitored through a variety of biomarkers—such as serum albumin, calcium, and Body Mass Index (BMI)—collected over time. A primary objective for clinicians is to understand how the trajectories of these biomarkers correlate with "time-to-event" outcomes, most notably mortality.
However, traditional statistical approaches often struggle with the complexities of real-world clinical data. Standard models frequently fail to account for the biological and temporal correlations between different biomarkers, leading to inefficient estimations. Furthermore, clinical data is often characterized by irregular sampling; while one biomarker might be measured monthly, another might only be recorded every six months. Traditional methods like "Last-Observation-Carried-Forward" (LOCF) introduce significant bias, and simple linear models are often too rigid to capture the non-linear fluctuations inherent in a patient's health status.
Methodology
To address these challenges, we introduce a flexible Bayesian framework for joint longitudinal-survival modeling. This framework is comprised of two integrated components designed to handle multiple biomarkers simultaneously while maintaining computational efficiency.
The Longitudinal Component
Rather than relying on simple linear regression, the longitudinal component utilizes a Multivariate Gaussian Process (GP). This approach allows the model to capture complex, non-linear trajectories across multiple data streams. To manage the relationship between different biomarkers and their progression over time, we employ a separable cross-covariance structure using Kronecker products. This mathematical formulation enables the model to account for the correlation between different biomarkers ($R$) and the temporal correlation ($S_i$) simultaneously.
One significant hurdle with Multivariate GPs is their computational complexity, typically $O(n^3)$. To make this framework practical for real-world applications, we implemented an eigenvalue decomposition technique that achieved a 30-fold speedup in computation, allowing for the analysis of larger patient cohorts.
The Survival Component
To avoid the restrictive assumptions of standard survival models, we utilize an infinite mixture of Weibull distributions via a Dirichlet Process Mixture (DPM). This allows the baseline hazard to take on a highly flexible shape, effectively clustering patients with similar baseline risk profiles rather than assuming a uniform hazard curve across the entire population.
Experimental Evaluation
Synthetic Data
We evaluated the model across three synthetic scenarios designed to test its robustness against varying levels of correlation, measurement frequencies, and biomarker volatility. In these simulations, the multivariate model consistently achieved a lower Mean Squared Error (MSE) in predicting missing biomarker values compared to univariate models. The performance gains were most significant when biomarkers were highly correlated ($ ho > 0.5$).
DaVita Hemodialysis Study
The framework was applied to a real-world cohort of 929 hemodialysis patients (2007–2011) from U.S. clinics. The analysis focused on the relationship between mortality and two key biomarkers: serum albumin and calcium.
By "borrowing strength" across data streams, the multivariate model utilized high-frequency measurements of one biomarker to improve the prediction of less frequently measured metrics. This resulted in a significant increase in precision; the multivariate model reduced the width of the 95% credible intervals for the estimated effects of albumin and calcium by 32% and 36%, respectively, compared to univariate approaches.
Key Insights and Clinical Findings
Our results demonstrate that traditional models, such as LOCF Cox regression, tend to shrink effect estimates toward the null. By accurately imputing biomarker values at the exact time of a survival event using the GP, our joint model captures the true magnitude of clinical risk.
Specifically, the model revealed that a 1 g/dL decrement in albumin is associated with a 5.1 times higher risk of death. This is a substantially stronger association than the 3.4 times risk estimated by standard LOCF Cox models, suggesting that traditional methods may significantly underestimate the mortality risk associated with declining biomarker levels.
Discussion and Future Work
While the current framework provides a robust and flexible solution for joint modeling, several avenues for improvement remain. Although the 30x speedup is substantial, the reliance on Markov Chain Monte Carlo (MCMC) sampling still poses scalability challenges as datasets grow into the tens of thousands of subjects. Future research will explore Variational Inference as a more scalable alternative for massive electronic health record (EHR) datasets.
Additionally, the current model operates under proportional hazards assumptions. Future iterations could relax this to allow for time-varying effects of the coefficients themselves, providing an even more granular understanding of how biomarker dynamics influence patient survival over time.
Summary updated