Dates
Monday, May 02, 2022 - 01:00pm to Monday, May 02, 2022 - 02:00pm
Location
NCS 220, or Zoom
Event Description

Abstract:
I present multiple computational models for extracting relevant features and learning predictive representations for neuroimaging and genetic data in UK-Biobank. These models are designed with the goal of providing improved diagnoses and a better understanding of medical conditions and other phenotypes. This includes a machine learning based predictive framework called Neuropredictome that identifies statistically significant linkages between 4928 phenotypes and neuroimaging features of 19,831 subjects. I also provide a novel quantitative method that uses deep learning based text embeddings to evaluate how well Neuropredictome's results align with 14,371 previously published peer-reviewed research articles. Next, I present a generalized framework based on state space systems that bridges the gap between network theory and control theory and extracts fMRI derived control circuits. This framework has the scalability required to mine mega-scale datasets, hitherto not possible using existing methods. In a purely data-driven manner, without priors, I demonstrate that the framework identifies thalamus-linked prefrontal-limbic and ventral stream subcircuits, selectively engaged during sensorimotor processing of affective and non-affective stimuli. I demonstrate that circuit-wide dysregulation, defined by degree of drift from healthy trajectories, tracks symptom severity for neuroticism, depression, and bipolar disorder. I also present methods for constructing low-dimensional vector representations (embeddings) of large-scale genotyping data, capable of reducing genotypes of hundreds of thousands of SNPs to 100-dimensional embeddings that retain substantial predictive power for inferring medical phenotypes. I also demonstrate how these genotype embeddings can be used for sharing sensitive medical data while preserving subject anonymity. Finally, using structural and functional neuroimaging in conjunction with cognitive tests, I show that type 2 diabetes mellitus accelerates brain aging and cognitive decline. Together, I believe such computational techniques can significantly advance modern medicine and treatment while enabling several scientific discoveries that revolutionize human health.

Reception to follow.

Contact events [at] cs.stonybrook.edu for Zoom information.

Event Title
Ph.D. Thesis Defense, Fahad Sultan: 'Predictive Models and Representations for Neuroimaging and Genetic Data'