Machine Learning for Healthcare 2019
University of Michigan, August 8-10, 2019
2019 Workshop - Community Data Challenge
Important: All participants must complete several steps prior to the Data Challenge —> instructions.
Thursday August 8th, 2019, University of Michigan North Quadrangle Dining Hall (105 S State St, Ann Arbor MI)
-----Session A -----
10:00am-12:00pm Learn. Introduction to the datasets and Google Cloud Platform (GCP)
What are the data? Where did they come from?
10:00 - 10:10 - Welcome - Brahmajee Nallamothu, MD & Jenna Wiens, PhD
10:10 - 10:30 - Responsible Use of Data - Erin Kaleba, Director, Data Office for Clinical and Translational Research, UM
10:30 - 10:50 - Introduction to the UM-MGI Dataset, Michael Sjoding, MD Assistant Professor of Pulmonary and Critical Care, UM
10:50 - 11:10 - Introduction to Philips Dataset, Omar Badawi, PharmD, MPH, Philips Healthcare
11:10 - 11:30 - Enabling reproducible research with PhysioNet, Tom Pollard, PhD, MIT
11:30 - 11:50 - Intro to GCP
11:50- 13:15 Lunch & Connect. Team Formation
-----Session B -----
13:15 - 14:15 Explore. Hands-on worksheet/tutorial exploring the data within GCP.
13:15 - 13:45 - MGI - Hands-on worksheet, John Vandervest, MS, UM and Shengpu Tang, UM
13:45 - 14:15 - Philips- Hands-on worksheet, Tom Pollard, Alistair Johnson, & Omar Badawi
14:15 - 16:00 Collaborate. Discuss opportunities with teammates.
Working in interdisciplinary teams articulate a technically interesting,
clinically relevant problem that could be solved using these data.
16:00 - 17:00 Pitch. Each team will present their problem to a panel of judge.
17:00 - 17:30 GCP Prizes & Closing Remarks.
Details: With an effort to engage interdisciplinary teams and accelerate research in precision health, we will introduce two datasets at this event i) the clinical data from the Michigan Genomics Initiative Cohort and ii) the Philips eICU dataset. These datasets will be hosted through Google Cloud Platform (GCP). Working in interdisciplinary teams, researchers will explore the data in GCP which has been tailored to working with health data. Once familiar with the dataset, teams will formulate technically interesting AND clinically relevant problems that can be solved using the data. Teams are invited to then pitch their ideas to a panel of experts. To encourage teams to continue working together to answer their research questions, winning teams will receive up to $1,000 in Google Cloud Platform credits (sponsored by Google).
2019 Main Conference Agenda
Friday August 9, 2019, University of Michigan Rackham Graduate School (915 E Washington St, Ann Arbor MI)
8:00- 8:45 Breakfast & Check-in
8:45 - 9:00 Welcoming Remarks
-----Session 1 -----
9:00- 9:45 Noémie Elhadad, PhD, Associate Professor in Biomedical Informatics and Computer Science, Columbia University
Data-Powered Women's Health: Endometriosis Characterization and Self-Management
Abstract: Endometriosis is a chronic, inflammatory, and estrogen-dependent condition with a high burden on quality of life, estimated to affect 6-10% of women of reproductive age worldwide. Despite its high prevalence, it is an enigmatic condition: there is currently no cure and no known biomarker or non-invasive diagnostic test for this multifactorial disease. In this talk I will report on ongoing research on two inter-related questions: how to characterize and discover the different ways in which endometriosis presents in individuals, essentially phenotyping the disease, and how to support individuals with self-discovery and management about the disease considering its heterogeneous presentations. I will show the current characterization of endometriosis from clinical data sources and discuss its current limitations, specifically the disconnect with the day-to-day patient experience of endometriosis. I will present the design and development of a personal health informatics solution (a research app called Phendo) and the analysis of the data contributed by Phendo participants towards phenotyping endometriosis. Finally, I will discuss how these data can be leveraged further to support individuals in learning about and self-managing their condition, as well as facilitating shared decision making with their providers.
9:45 - 10:15 - Contributed Spotlights - Research Paper Track
10:15 - 10:45 Coffee Break and Discussion
10:45-11:30 Andrew Rosenberg, MD, Chief Information Officer for Michigan Medicine, University of Michigan
Supporting Advanced Analytics at Michigan Medicine: Experiences and Lessons Implementing Novel Data and Information Solutions at UM
Abstract: The CIO for the University of Michigan Health System and Medical School will discuss trends in a variety of analytic techniques, and examples where they have been trialed at Michigan Medicine. Often, variable faculty expertise, maturity of the analytic solutions, and quality of source data initially yield suboptimal results for attempting to trial novel analytic techniques at the scale of a modern health system. However, several reproducible approaches to solving complex analytic questions have been established and will be discussed in the context of specific examples at the University of Michigan. Among these include broad knowledge of the data sources, repositories and analytic user groups where expertise reside, competing requests that are harmonized with each other, and core IT/IS support combined with more unique knowledge and innovative approaches from faculty. We will discuss several examples deploying a variety of machine learning approaches, use of enterprise applications’ advanced analytics capabilities such as Epic Cognitive Computing etc. We will also highlight how these have been performed in light of expanding cyber-security and information assurance rigor across the health system, medical school and university.
11:30 - 12:00 - Contributed Spotlights - Research Paper Track
12:00-14:00 Posters A & Lunch
-----Session 2 -----
14:00-14:45 Pilar N Ossorio, PhD, JD, Professor of Law and Bioethics, Law School, University of Wisconsin
Justice in Machine Learning for Health Care
Abstract: Understanding when an inequality in health or health care reflects injustice is more difficult than understanding injustice in areas such as hiring or school admissions. For instance, if fairness means treating people alike to the extent their relevant characteristics are the same, then we need to agree on what characteristics are relevant. In hiring decisions, we might say that it is unfair to treat people with the same or similar job qualifications differently based on their race, because qualifications are relevant and race is not. In health and healthcare, we often do not know whether gender, age, class, race, or other characteristics are relevant to people’s diagnoses, treatment responses, prognoses, or health risks. A wealth of data indicates that pernicious stereotypes and social inequalities influence health and health care, and no doubt these social problems will pattern the data on which algorithms are trained. On the other hand, not all gender, class, racial, or other differences in health or health care reflect injustice. This presentation will develop a taxonomy of ways in which health and health care inequalities arise; discuss how and why inequalities count as injustices; and discuss ways in which machine learning could help ameliorate injustice in health care, or contribute to it.
14:45 - 15:15 - Contributed Spotlights - Research Paper Track
15:15 - 15:45 Coffee Break and Discussion
15:45-16:30 Mert Sabuncu, PhD, Assistant Professor, School of Electrical and Computer Engineering, Cornell University
Adaptive Compressed Sensing MRI with End-to-End Deep Learning
Abstract: Magnetic Resonance Imaging (MRI) can be accelerated by sampling below the Nyquist rate. In this talk, I will consider the problem of optimizing the under-sampling pattern in a data-driven fashion. For a given sparsity constraint, our method optimizes the under-sampling pattern and reconstruction model, using an end-to-end learning strategy. Our algorithm learns from full-resolution data that are under-sampled retrospectively. The proposed method, which we call LOUPE (Learning-based Optimization of the Under-sampling PattErn), was implemented by modifying a U-Net, a widely used convolutional neural network architecture, that we append with the forward model that encodes the under-sampling process. Our experiments with brain and knee MRI scans show that the optimized under-sampling pattern can yield significantly more accurate reconstructions compared to standard under-sampling schemes.
16:30 - 17:00 - Contributed Spotlights - Research Paper Track
17:00 - Posters B & Reception
Saturday August 10, 2019
-----Session 3 -----
9:00-9:45 Mihaela van der Schaar, PhD, John Humphrey Plummer Professor of ML, AI and Medicine, University of Cambridge
Learning Engines for Healthcare: Transforming Medicine through AI-enabled Healthcare Pathways
Abstract:In this talk, I will discuss recent machine learning and AI theory, methods, algorithms and systems which we developed in our lab to understand the basis of health and disease, to catalyze clinical research, to support clinical decisions through individualized medicine, to inform clinical pathways, to better utilize resources & reduce costs and to inform public health.
To do this, we are creating what I call Learning Engines for Healthcare (LEH’s). An LEH is an integrated ecosystem that uses machine learning, AI and operations research to provide clinical insights and healthcare intelligence to all the stakeholders (patients, clinicians, hospitals, administrators). In contrast to an Electronic Health Record, which provides a static, passive, isolated display of information, an LEH provides dynamic, active, holistic & individualized display of information including alerts.
In this talk, I will focus on 3 steps in the development of LEH’s: 1. Building a comprehensive model that accommodates irregularly sampled, temporally correlated, informatively censored and non-stationary processes in order to understand and predict the longitudinal trajectories of diseases. 2. Establishing the theoretical limits of causal inference and using what has been established to create a new approach that makes it possible to better estimate individualized treatment effects. 3. Using Machine Learning itself to automate the design and construction of entire pipelines of Machine Learning algorithms for risk prediction, screening, diagnosis and prognosis.
9:45-10:15 Contributed Spotlights - Clinical Abstract Track
10:15-10:45 Coffee Break and Discussion
10:45-11:30 Roy Perlis, MD, MSc, Professor of Psychiatry, Harvard University
Controversies in Diagnosis and Treatment of Major Depressive Disorder
Abstract: The impact of major depressive disorder is profound, whether measured in terms of morbidity and mortality, health care costs, or impact on productivity and quality of life. Efforts to improve the diagnosis and treatment of this brain disease have been hindered by a lack of understanding of etiology, overlap with other diseases, and optimal treatment strategies. This presentation will focus on emerging ideas about major depressive disorder as they inform efforts to apply machine learning to improve diagnosis, select optimal treatments, and stratify risk for adverse outcomes, including suicide. More broadly, it will introduce concepts relevant to the use of large clinical data sets where both phenotypes and outcomes may be difficult to reliably characterize.
11:30-12:00 Contributed Spotlights - Clinical Abstract Track
12:00-13:30 Lunch & Clinical Track Posters
-----Session 4 -----
13:30-14:15 Rajan Dewar, MD, PhD, Associate Professor, Michigan Medicine, University of Michigan
ML in Global Health: Can ML Tools Help Implement Cost Effective Healthcare Solutions in Low and Middle Income Countries?
Abstract: Health care delivery in LMICs are impeded by the lack of high quality and more expensive tools that enables precise diagnosis and treatment, which is available and affordable in western countries. We propose that ML tools can help bridge this lack of resources. Pathology infrastructure is necessary for early (screening) diagnosis, appropriate diagnosis and treatment selection of cancer which is increasing in LMICs. Cervical cancer is one of the leading causes of cancer death in women. Screening at early stages using the popular Pap smear test has been demonstrated to reduce fatalities significantly. Cost effective, automated screening methods can significantly improve the adoption of these tests worldwide. Automated screening involves image analysis of cervical cells. Gaussian Mixture Models (GMM) are widely used in image processing for segmentation which is a crucial step in image analysis. In our proposed method, GMM is implemented to segment cell regions to identify cellular features such as nucleus, cytoplasm while addressing shortcomings of existing methods. This method is combined with shape based identification of nucleus to increase the accuracy of nucleus segmentation. This enables the algorithm to accurately trace the cells and nucleus contours from the pap smear images that contain cell clusters. The method also accounts for inconsistent staining, if any. The results that are presented shows that our proposed method performs well even in challenging conditions. Despite this, challenges remain and will be discussed. Other population health ML solutions include - high resolution databases and survey tools that can identify at-risk populations that can help targeted screening.
14:15 - 15:00 Katherine Heller, PhD, Duke University and Google Medical Brain
Machine Learning for Healthcare in Practice
15:00-16:00 Panel Discussion, moderated by David Kale, PhD
Omar Badawi, PharmD, MPH, Philips Healthcare
Luca Foschini, PhD Co-founder Evidation
Katherine Heller, PhD, Google
Sachin Kheterpal, MD, MBA University of Michigan
Pilar N. Ossorio, PhD, JD, University of Wisconsin
16:00-16:15 Closing Remarks
16:30 Feedback Discussion Session