Raghav Tandon - Machine Learning PhD Student - Department of Biomedical Engineering (BME)
Date: June 26th
Time: 10:00 AM – 12:00 PM ET
Location: Virtual
Meeting Link: https://gatech.zoom.us/j/93553577940?pwd=ZlZLRmFwZ1VNNjVHOVJDQnd1TURYUT09
Committee
- Dr. Cassie S Mitchell (Advisor, BME, GT)
- Dr. Saurabh Sinha (BME and ISyE, GT)
- Dr. Yajun Mei (ISyE, GT)
- Dr. James J Lah (Department of Neurology, Emory School of Medicine)
- Dr. Greg Gibson (School of Biological Sciences, GT)
Abstract
Alzheimer's disease (AD) is the most prevalent form of dementia. It has a multifactorial, heterogeneous presentation where pathological changes can occur decades before symptoms appear. The objective of this work was to develop novel machine learning algorithms that enable dynamic, multimodal stratification of patients using cross-sectional data. The end goal was to employ artificial intelligence to identify high-risk patients while still in the functionally unimpaired state. First, supervised backward feature selection approaches were employed to identify new protein biomarkers for AD. The identified markers were validated across cohorts and showed a functional enrichment in sugar metabolism. Second, a new probabilistic generative algorithm was developed to learn disease progression trajectory from cross-sectional proteomic and imaging data. Disease progression was defined by an unobserved sequence of biomarker abnormalities, which is inferred by the model. Predecessors of the model have faced computational complexities due to the space of possible sequences increasing factorially with the number of included data features. The newly developed scaled event-based model (sEBM) addressed these computational challenges to enable stratification of disease progression risks while adjusting for demographical and genotype covariates. Finally, sEBM was further extended to infer patient heterogeneities (or subtypes) in disease progression. The data likelihood was modeled as a mixture of multiple disease progression trajectories which were jointly optimized within an expectation-maximization framework. The inferred trajectories corresponded to different AD subtypes, and the positions along a trajectory corresponded to progressive AD stages. The inferred AD subtypes captured differences in demographics (age, gender), medical history, age at symptom onset and levels of key disease markers. The inferred disease stages captured differences in neuropsychological test scores, brain region volumes, and risk of future conversion to AD. Future clinical translation of the developed algorithms will improve personalized medicine and clinical decision making by enabling real-time, state-of-the-art asymptomatic patient risk stratification.