CHAI Seminar Series

Series Type
Event Series
Associated Organization(s)
Associated Organization(s)

Publication Search Results

Now showing 1 - 9 of 9
  • Item
    How Do We Accelerate Data Driven Health Care?
    ( 2018-04-19) Stewart, Walter
    By driving the digitization of healthcare data, the 2009 HITECH Act set the stage for business and business practice transformation. Today, the amount of healthcare data doubles every two years and growth is accelerating. It is now widely accepted that the future of healthcare will strongly depend on the effective and maximal use of the growing diversity and volume of healthcare data to transform care in ways that substantially improve patient accessibility, quality, and affordability of care and patient and provider experience. But, achieving that vision – transforming healthcare through the continual creation of analytical insights and their translation into clinical practice – is exceedingly difficult in the complex business of health care because it requires diverse knowledge, talent, assets, and endless financial investment. No single entity or market has all that is required. Today, these capabilities and assets rest within distinctly different organizations that largely act independently of each other. In this talk I will describe an organizational model that bridges the divides among organizations to accelerate transformation of health care through data driven science, methods, and solutions.
  • Item
    Translating from Chemistry to Clinic with Deep Learning: Modeling the Metabolism and Subsequent Reactivity of Drugs
    ( 2018-04-12) Swamidass, S. Joshua
    Many medicines become toxic only after bioactivation by metabolizing enzymes. Often, metabolic enzymes transformed them into chemically reactive species, which subsequently conjugate to proteins and cause adverse events. For example, carbamazepine is epoxidized by P450 enzymes in the liver, but then conjugates to proteins, causing Steven Johnsons Syndrome in some patients. The most difficult to predict drug reactions, idiosyncratic adverse drug reactions (IADRs), often depend on bioactivation. Our group has been using deep learning to model the metabolism of diverse chemicals, and the subsequent reactivity of their metabolites. Deep learning systematically summarizes the information from thousands of publications into quantitative models of bioactivation, modeling precisely how medicines are modified by metabolic enzymes. These models are giving deeper understanding of why some drugs become toxic, and others do not. At the same time, deep learning can be used to understand drug toxicity as it arises in clinical data, and why some patients are affected, but not others. A conversation between the basic and clinical sciences is now possible, where patient outcomes can be understood in light of bioactivation mechanisms, and these mechanisms can explain why some patients are susceptible to drug toxicity, and others are not.
  • Item
    Causal Network Discovery from Biomedical and Clinical Data
    ( 2018-04-05) Cooper, Gregory
    This talk will provide an introduction to concepts and methods for learning causal relationships in the form of causal networks from biomedical and clinical data, including solely observational data. Examples will be given of applying these methods to biomedical data. The talk will also provide pointers to software for learning causal networks from data, including data containing thousands of variables.
  • Item
    Big Data in Behavioral Medicine
    ( 2018-03-06) Rehg, James M.
    The explosion of health-related data in the form of electronic health records and genomics has captured the attention of both the machine learning and medical informatics communities. In his talk, Dr. Rehg will describe an emerging opportunity to bring data analytics to bear on the behavioral dimension of health. Among other things, he will discuss advances in mobile sensing, from classical activity monitoring to the recent advent of wearable cameras, and how these rapidly evolving technologies provide new opportunities to continuously-measure behaviors under naturalistic conditions and construct novel predictive models for adverse behavioral outcomes.
  • Item
    Machine Learning from Irregularly-Sampled Temporal Data: A Case Study in Predicting Across Most Diseases in Electronic Health Records
    ( 2018-02-27) Page, David
    Much of the world’s real data on people is irregularly-sampled, temporal, and observational (meaning we don’t get to experiment as in a randomized clinical trial). For example, customers make purchases on various dates of their choice, not necessarily once a week or once a month, and we only observe rather than intervene in their decisions. Patients visit the doctor whenever they feel the need, and we observe their doctors’ entries in the electronic health record (EHR), without the ability to randomize patient treatments. We show that despite this lack of control or sampling regularity, we can predict future events from such data with surprising accuracy, for example better than 80% on average across a variety of diagnosis codes in the EHR a month in advance. We further show that despite many types of potential confounding, we can actually discover causal factors (e.g., effect of a drug on a disease or on a measurement such as blood pressure) at similar levels of accuracy for real problems. The key to doing so is modeling person-specific, time-varying baseline levels, e.g. of a measurement such as blood pressure or a risk such as for heart attack. On the applied side this talk will focus entirely on medical applications, but the approaches developed and employed are general-purpose machine learning algorithms with broad potential applicability.
  • Item
    Unsupervised Phenotyping using Tensor Factorization
    ( 2018-02-20) Perros, Ioakeim
    How can we distill multiple aspects of raw and noisy electronic health record data, such as diagnoses, medications and procedures, to a few concise clinical states without human-annotated labels? In this talk, we present a review of works tackling this challenge through the use of tensor factorization methods. Several aspects of this problem will be discussed such as scalable and efficient computations, interpretability of the results and handling temporally-evolving data.
  • Item
    Predicting Hospital Readmissions Among Kidney Transplant Recipients
    ( 2018-02-13) Patzer, Rachel
    More than half of kidney transplant recipients are rehospitalized in the year after kidney transplantation, placing a large burden on the hospital system and placing patients at higher risk for poor health outcomes, including infection, graft failure and/or death. Several studies have identified risk factors for hospital admission using traditional statistical methods, but models have been limited by moderate predictive accuracy, the use of static (rather than dynamic) models, and the use of administrative data that may not capture the changing risk factors pre- to post-transplant. The overall goal of this research is to develop and validate rigorous, dynamic risk prediction models, integrate these models within a clinician dashboard to allow for real-time decision making and appropriate interventions for high risk patients. This presentation will discuss the importance of the clinical problem, as well as preliminary data analyses from national surveillance data models and local Emory Transplant Center models. The long term objective of this project is to improve health outcomes of patients and improve the efficiency of clinical care for transplant recipients.
  • Item
    Precision Medicine at Georgia Tech: Introduction to the Health Data Analytics Platform
    (Georgia Institute of Technology, 2018-02-06) Duke, Jon
    Conducting health research is complex, from accessing health data to running analytics to building interoperable applications. The Georgia Tech Health Data Analytics Platform (HDAP) is designed to expedite the process of health research through 1) accessible datasets, 2) analytics tools, and 3) interoperable FHIR APIs and and a FHIR application deployment environment. In this presentation, we will discuss precision medicine research, the use of natural language processing, and the HDAP platform.
  • Item
    Using Electronic Health Records to Support Patient Care and Clinical Research
    ( 2018-01-30) Weng, Chunhua
    Electronic health records (EHR) contains rich clinical phenotype information. In this talk, I will present methods and early results from two projects to demonstrate the potential of using EHR data to facilitate precision medicine and optimize clinical research towards a learning health system. In project one, we developed a phenotype-driven diagnostic decision support system, where Human Phenotype Ontology (HPO) concepts were extracted from EHR narratives and used to prioritize disease genes based on the HPO-coded phenotypic manifestations. We tested this approach on 28 pediatric patients with confirmed diagnoses of monogenic diseases, and found that the causal genes were ranked among the top 100 genes out of > 25000 genes for 16/28 cases (P<2.2x10-16), demonstrating the promise of leveraging EHR data to automate phenotype-driven analysis of clinical exomes or genomes and implement genomic medicine on scale. In project two, we developed a metric called GIST, which stands for The Generalizability Index of Study Traits, to assess the population representativeness of clinical trials by using EHR data to profile the target populations for clinical trials and by comparing the study populations to the target populations. GIST enables us to improve the transparency of population representativeness of clinical studies and to help clinical researchers to make informed decisions to optimize patient selection.