Bayesian blockwise inference for joint models of longitudinal and multistate processes

TL;DR

This paper introduces scalable Bayesian blockwise inference methods for joint models of longitudinal and multistate processes, enabling analysis of complex event data in large biomedical datasets with improved computational efficiency.

Contribution

The paper develops two novel Bayesian inference approaches for joint multistate models, facilitating analysis of complex event processes and large datasets with parallel computing and variable selection.

Findings

Methods achieve accurate posterior estimates and improved sampling efficiency.

Application to UK health records reveals distinct associations between blood pressure and disease transitions.

Approaches are scalable and straightforward to implement for large biomedical data.

Abstract

Joint models (JM) for longitudinal and survival data have gained increasing interest and found applications in a wide range of clinical and biomedical settings. These models facilitate the understanding of the relationship between outcomes and enable individualized predictions. In many applications, more complex event processes arise, necessitating joint longitudinal and multistate models. However, their practical application can be hindered by computational challenges due to increased model complexity and large sample sizes. Motivated by a longitudinal multimorbidity analysis of large UK health records, we have developed a scalable Bayesian methodology for such joint multistate models that is capable of handling complex event processes and large datasets, with straightforward implementation. We propose two blockwise inference approaches for different inferential purposes based on different levels of decomposition of the multistate processes. These approaches leverage parallel computing, ease the specification of different models for different transitions, and model/variable selection can be performed within a Bayesian framework using Bayesian leave-one-out cross-validation. Using a simulation study, we show that the proposed approaches achieve satisfactory performance regarding posterior point and interval estimation, with notable gains in sampling efficiency compared to the standard estimation strategy. We illustrate our approaches using a large UK electronic health record dataset where we analysed the coevolution of routinely measured systolic blood pressure (SBP) and the progression of multimorbidity, defined as the combinations of three chronic conditions. Our analysis identified distinct association structures between SBP and different disease transitions.