Time-Dynamic Tree-Based Methods for Personalized Alzheimer's Disease Prediction

Project Summary (Abstract) This two-year grant proposal responds to PAR-23-179, and its long-term goal is to advance the understanding of biomarker in relation to preclinical Alzheimer’s Disease (AD). This proposal seeks to leverage interpretable and flexible tree-based methods, clinically and biologically informed tree decision rules, longitudinal

biomarkers, and risk and protective factors, to develop novel statistical methods and construct tree-based algorithms. The novel methods and tree-based algorithms will help identify time-dynamic and personalized biomarker subgroups at high risk for cognitive decline due to AD and predict progression risks.

AD is a devastating disease affecting over 6 million people in the U.S. and has burdened the U.S. healthcare system and caregivers with increases. Importantly, evidence suggests that the pathophysiological process begins many years, if not decades, before the diagnosis of AD dementia, and recent findings demonstrate that

biomarker deterioration starts many years before cognitive decline due to AD. Identifying high-risk subgroups based on biomarker information will expand the window of opportunity during which therapeutic intervention may have the greatest potential for success. Tree-based methods appear well-suited for producing clinically applicable decision rules that leverage

complicated interactions between different biomarkers, and between biomarkers and risk factors. But existing tree-based methods are limited in clinical relevance, biological interpretability, and statistical inference, and novel methods are needed. Meanwhile, the consortium of multiple longitudinal follow-up studies presents new

opportunities and challenges. The Preclinical AD Consortium (PAC) data comprises five studies that have been collecting longitudinal clinical, cognitive, biomarker, and genetic data from individuals who were cognitively normal when first enrolled and followed for many years. The large sample size and the breadth of

the merged and harmonized data create opportunities for more precise and personalized classification and risk prediction, based on longitudinal biomarkers and risk and protective factors. These opportunities also create challenges for method development to be time-dynamic and personalized. Projects supported by this proposal will seek to develop novel tree-based statistical methods for subgroup

identification and risk prediction for cognitive decline due to AD, and construct classification and prediction algorithms using the PAC data. Our research team will pursue these goals through two Specific Aims: (1) establish a time-dynamic and personalized statistical classification and prediction framework using tree-based

methods; and (2) leverage the PAC data to construct classification and prediction algorithms for onset of AD- related clinical symptoms. Successful completion of these Aims will produce novel tree-based methodologies and generate tree-based classification and prediction algorithms that predict onset of AD-related symptoms.