Bayesian shape-constrained regression for quantifying Alzheimer's disease biomarker progression

Abstract

Several biomarkers are hypothesized to indicate early stages of Alzheimer's disease, well before the cognitive symptoms manifest. Their precise relations to the disease progression, however, is poorly understood. This lack of understanding limits our ability to diagnose the disease and intervene effectively at early stages. To provide better understanding of the relation between the disease and biomarker progressions, we propose a novel modeling approach to quantify the biomarkers' trajectories as functions of age. Building on monotone regression splines, we introduce two additional shape constraints to incorporate structures informed by the current medical literature. First, we impose the regression curves to satisfy a vanishing derivative condition, reflecting the observation that changes in biomarkers generally plateau at early and late stages of the disease. Second, we enforce the regression curves to have a unique inflection point, which enhances interpretability of the estimated disease progression and facilitates assessment of temporal ordering among the biomarkers. We fit our shape-constrained regression model under Bayesian framework to take advantage of its ability to account for the heterogeneity in disease progression among individuals. When applied to the BIOCARD data, the model is able to capture asymmetry in the biomarkers' progressions while maintaining interpretability, yielding estimates of the curves with temporal ordering consistent with the existing scientific hypotheses.