Abstract
Identifying reliable biomarkers for predicting clinical events in
longitudinal studies is important for accurate disease prognosis and the
development of new treatments. However, prognostic studies are often not
randomized, making it difficult to account for patient heterogeneity. In
amyotrophic lateral sclerosis (ALS), factors such as age, site of disease onset
and genetics impact both survival duration and biomarker levels, yet their
impact on the prognostic accuracy of biomarkers over different time horizons
remains unclear. While existing methods for time-dependent receiver operating
characteristic (ROC) analysis have been adapted for censored time-to-event
outcomes, most do not adjust for patient covariates. To address this, we
propose the nonparanormal prognostic biomarker (NPB) framework, which models
the joint dependence between biomarker and event time distributions while
accounting for covariates. This provides covariate-specific ROC curves which
assess a potential biomarker's accuracy for a given time horizon. We apply this
framework to evaluate serum neurofilament light (NfL) as a biomarker in ALS and
demonstrate that its prognostic accuracy varies over time and across patient
subgroups. The NPB framework is broadly applicable to other conditions and has
the potential to improve clinical trial efficiency by refining patient
stratification and reducing sample size requirements.