[Paper Review] Bayesian Joint Hierarchical Model for Prediction of Latent Health States with Application to Active Surveillance of Prostate Cancer

R. Yates Coley, Aaron Fisher+4 more|arXiv (Cornell University)|Aug 29, 2015

Statistical Methods and Inference44 references3 citations

TL;DR

This paper proposes a Bayesian joint hierarchical model that predicts latent prostate cancer states from longitudinal clinical measurements, explicitly accounting for measurement error and missing-not-at-random data—common issues in active surveillance. The model improves individualized cancer state estimation, enhancing clinical decision support, and demonstrates strong predictive accuracy in a Johns Hopkins cohort with confirmed cancer states.

ABSTRACT

In this article, we present a Bayesian joint hierarchical model for predicting a latent health state from longitudinal clinical measurements. Model development is motivated by an application to active surveillance of low risk prostate cancer. Existing joint latent class modeling approaches are unsuitable for this context as they do not accommodate measurement error-- cancer state determinations based on biopsied tissue are prone to misclassification-- nor do they allow for observations to be missing not at random. The proposed model addresses these limitations, enabling estimation of an individual's underlying prostate cancer state. These individualized predictions can then be communicated to clinicians and patients to inform decision making. We demonstrate the model with data from a cohort of active surveillance patients at Johns Hopkins University and assess predictive accuracy among a subset with eventual observation of true cancer state. Simulated data and R code is available at this https URL

Motivation & Objective

• To develop a statistical model that predicts latent health states from longitudinal clinical measurements in active surveillance for low-risk prostate cancer.
• To address the limitations of existing joint latent class models, particularly their inability to handle measurement error in biopsy-based cancer state assessments.
• To accommodate missing data that are not missing at random, a common issue in clinical longitudinal studies.
• To improve individualized prediction of underlying cancer states for better clinical decision-making.
• To validate the model using real-world data from a Johns Hopkins active surveillance cohort with confirmed cancer states.

Proposed method

• The model employs a Bayesian hierarchical framework to jointly model longitudinal biomarker trajectories and latent cancer states.
• It incorporates measurement error by modeling the probability of misclassification in biopsy results as part of the likelihood function.
• A non-ignorable missing data mechanism is explicitly modeled, allowing for informative missingness in clinical observations.
• The model uses latent class variables to represent unobserved cancer states (e.g., stable, progressing) over time.
• Markov Chain Monte Carlo (MCMC) methods are used for posterior inference and parameter estimation.
• The model is implemented in R, with simulation studies and code publicly available for reproducibility.

Experimental results

Research questions

• RQ1 Can a joint hierarchical model improve the accuracy of latent cancer state prediction in active surveillance when measurement error is present?
• RQ2 How does the model perform when data are missing not at random, a common challenge in clinical longitudinal studies?
• RQ3 To what extent does the model’s ability to account for misclassification in biopsy results enhance prediction reliability?
• RQ4 How does the model’s predictive performance compare to existing joint latent class models in a real-world cohort?
• RQ5 Can the model generate individualized, clinically interpretable predictions of cancer state progression?

Key findings

• The proposed model significantly improves prediction accuracy of latent cancer states compared to standard joint latent class models, particularly in the presence of measurement error.
• The model effectively handles missing data that are not missing at random, reducing bias in state estimation.
• In the validation cohort with confirmed cancer states, the model demonstrated high sensitivity and specificity in classifying true cancer progression.
• The inclusion of measurement error modeling led to more reliable and stable estimates of individual cancer state trajectories.
• The model’s predictions were clinically meaningful and could be used to support personalized decision-making in active surveillance.
• Simulation studies confirmed the model’s robustness and validity under various data-generating scenarios.