Retarded kernels for longitudinal survival analysis and dynamic prediction

TL;DR

This paper introduces a retarded kernel method for dynamic survival prediction that balances complexity and data utilization, offering a practical alternative to joint models and landmark analysis in clinical settings.

Contribution

The paper proposes a novel retarded kernel approach that models hazard rates based on the full history of longitudinal covariates, bridging the gap between existing methods.

Findings

Predictive accuracy comparable to joint models and landmark analysis.

Model reduces to Cox model for time-independent covariates.

Two natural kernel parameterisations derived and tested.

Abstract

Predicting patient survival probabilities based on observed covariates is an important assessment in clinical practice. These patient-specific covariates are often measured over multiple follow-up appointments. It is then of interest to predict survival based on the history of these longitudinal measurements, and to update predictions as more observations become available. The standard approaches to these so-called `dynamic prediction' assessments are joint models and landmark analysis. Joint models involve high-dimensional parametrisations, and their computational complexity often prohibits including multiple longitudinal covariates. Landmark analysis is simpler, but discards a proportion of the available data at each `landmark time'. In this work we propose a `retarded kernel' approach to dynamic prediction that sits somewhere in between the two standard methods in terms of complexity. By conditioning hazard rates directly on the covariate measurements over the observation time frame, we define a model that takes into account the full history of covariate measurements but is more practical and parsimonious than joint modelling. Time-dependent association kernels describe the impact of covariate changes at earlier times on the patient's hazard rate at later times. Under the constraints that our model (i) reduces to the standard Cox model for time-independent covariates, and (ii) contains the instantaneous Cox model as a special case, we derive two natural kernel parameterisations. Upon application to three clinical data sets, we find that the predictive accuracy of the retarded kernel approach is comparable to that of the two existing standard methods.