Statistical Methods for Early Disease Prediction and Treatment Strategy Estimation Using Biomarker Signatures

点击翻译按钮获取中文摘要

Neuropsychiatric disorders pose an immense burden on patients, families, and health care systems, thus underscoring the urgent need to develop disease-modifying treatment. Research on neuropsychiatric disorders (e.g., Alzheimer’s disease, Parkinson’s disease) faces unique challenges, including the fact that these disorders typically have a late onset and slow progression, the diagnostic criteria are based on subjective clinical symptoms, and there is substantial disease and subject heterogeneity. In the proposed work, we aim to tackle these challenges by leveraging complementary contributions from multiple biomarkers, including genome-wide polymorphisms, whole brain neuroimaging, biofluids, and comprehensive neuropsychiatric assessments. We develop sophisticated analytic tools with higher resolution and improved accuracy by accounting for biological mechanisms of disease, synthesizing dynamic system-wide information, and integrating multiple sources of biomarkers. These methods are applied to clinical data collected by the investigative team or available from large international consortia in order to model the earliest pathological changes of neurodegenerative disease, assess treatment responses, and inform the design of early-intervention clinical trials and the discovery of optimal personalized therapies. Specifically, in Aim 1, we develop efficient methods for multi-level semiparametric transformation models to estimate and test the risk of genetic variants on various types of complex phenotypes to inform genetic counseling and improve clinical trial efficiency. Our methods do not rely on full pedigree genotyping and provide family-specific substructure, in addition to population substructure, to better control confounding and reduce false discovery rates in genome-wide association studies. In Aim 2, we develop large-scale nonlinear dynamic systems through ordinary differential equations with random inflections to understand early pathological changes and identify subjects with preclinical signs. Our method provides multi-domain integration of ensembles of biomarker dynamics. In Aim 3, we develop dynamic hazards models and incorporate dynamic network structures to estimate biomarker profiles that evolve smoothly with disease progression for earlier disease diagnosis. We account for irregularly measured biomarkers and biological network dependence among biomarkers. In Aim 4, we develop doubly robust and efficient machine learning methods to identify predictive markers, estimate optimal individualized therapies, and identify subgroups who may receive the greatest benefit from therapy, with minimal risk. In each aim, we will validate the proposed methods through extensive simulation studies and demonstrate their practical value via application to real-world clinical studies. We establish theoretical properties of the proposed methods using modern empirical process theory and statistical learning theory. Together, the state-of-the-art analytic methods proposed here will substantially improve analytic accuracy, and our combined statistical and clinical expertise will ensure that our methods are translated directly back to the clinical and translational research community.