Landmarking with Latent Class Mixed Models for Dynamic Prediction of Time-to-event Data with Heterogeneous Biomarker Trajectories

Abstract

The increasing ability to securely access electronic health records (EHR) has created unprecedented opportunities to monitor the health trajectories of large heterogeneous patient populations throughout their lifetime. Repeated measurements of time-varying covariates (such as biomarkers measured via routine blood tests) can inform dynamic risk prediction of time-to-event outcomes, updating risk estimates as new information becomes available. Existing dynamic risk prediction approaches often assume homogeneous longitudinal trajectories across individuals. This assumption is not met when there is heterogeneity driven by latent subgroup structure (e.g. due to unobserved confounders), as is often the case with real-world biomedical data. At present, accounting for such heterogeneity is only available in joint latent class models for longitudinal and time-to-event data, but they are computationally intensive, often prohibitively so for large-scale data, such as those present in EHR settings. To address these challenges, we propose a novel landmarking approach that integrates latent class mixed models (LCMMs) to capture latent heterogeneity in longitudinal trajectories. Our method is implemented in a modular R package, landmaRk, which is available on CRAN and allows users to flexibly specify the components of a landmarking analysis, beyond our proposed approach. Through simulation studies, we demonstrate improvements in prediction performance in the presence of latent heterogeneity compared to traditional landmarking strategies, while remaining computationally efficient for large datasets. We also provide a proof-of-concept illustration using real data.