Data-driven Subtypes of Alzheimer's disease progression for targeted treatment

Alzheimer's disease (AD) is the only of the top ten leading causes of death that cannot be cured. AD drug development has been impeded by significant disease heterogeneity. Despite the massive information that we have been collecting from patients in the last decade, there is still a lack of understanding of why some AD patients are fast progressors and others are slow progressors; such gaps in knowledge have posed many barriers to conducting targeted clinical trials or developing effective personalized therapies. If we can select more homogenous patients (based on neuropathology, biomarkers, demographics, and clinical presentations) who might benefit from a specific intervention, more focused trials would then be made possible to accelerate new therapy development. However, identifying homogeneous patient subpopulations is a non-trivial problem. Most of the existing research strategies lack a longitudinal and holistic consideration of multi-modal, multi-resolution, and multi-source data. We believe there are unique opportunities for us to address this challenge with advanced machine learning by integrating heterogeneous information. Clinical trials are the de facto gold standard for monitoring the progression of AD. Our team has access to 14 AD randomized trials data, as well as three observational trials and registry data. An innovative exploration of big data through advanced informatics models will address the heterogeneity of AD. Our goal is to develop novel machine learning models to reveal and stratify heterogeneous subpopulations based on risk and progression patterns. We will transform them into potentially targetable groups to enable focused trials, support future therapeutic development, and promote personalized AD treatment. Specifically, we will develop novel deep temporal clustering models to identify subtypes of progression in AD patients using brain atrophy, fluid biomarkers, and cognitive decline (Aim 1). We will also develop a scalable casual structure model to clinical trajectory to AD (Aim 2). Finally, we will verify and connect AD subpopulation identified in Aim 1 and 2 to targetable population.