Noah Black
Pronouns: He/Him
Research Mentor(s): Sriram Chandrasekaran
Research Mentor School/College/Department: Biomedical Engineering / Medicine
Program: UROPF
Session: Session 2 (10:00am – 10:50am)
Authors: Noah Black, Awanti Sambarey, Kirk Smith, Sriram Chandrasekaran
Abstract
Tuberculosis is a globally substantial disease. With over ten million diagnoses in 2015 and over one million deaths, it is clear tuberculosis is an unresolved global health concern. Unfortunately, while tuberculosis rates have been decreasing globally, there has been a substantial increase in the presence of antibiotic resistant tuberculosis. Therefore, the National Institutes of Health has initiated the TB Portals effort, which is an endeavor to collect data about tuberculosis. TB Portals agglomerates data from numerous countries in eastern Europe, Asia, and Africa. Our goal is to design machine learning models that predict patient outcomes by using TB Portal’s data. This effort aims to better understand what factors relate to patient outcome as well as provide physicians an understanding of at-risk patients. Thus far, we have implemented a state-of-the-art random forest model using patient social data that increases model performance across every metric we have tested. Additionally, our efforts of developing cohort specific models have provided a hypothesis for how patient outcome variance is captured by data modalities. For instance, for drug sensitive tuberculosis, experiments suggest very little outcome variance is captured by recorded social variables. Finally, our future efforts will incorporate direct and automatic analysis of patient anatomical data (CXR images) through deep learning and model ensembling. These experiments will attempt to compress outcome specific image features such that they can be concatenated with social data without hindering model performance.
