Date: Friday, July 27th
Time: 4:30 – 6:00pm
At the AI in Healthcare session of the CIFAR Deep Learning Summer School we will discuss the potential of AI to have real-world impact in healthcare, whether that means improving diagnostic accuracy, quality of life improvements for patients, or long-term preventative care. After introductory talks in the main lecture hall, we will split up into ten interactive sessions, each focusing on a different aspect of healthcare and clinical data and led by a leader in health research. At these breakout sessions, attendees will be asked to participate in a brainstorming session to identify ways in which their machine learning expertise could be applied to the healthcare challenged being discussed. Finally, we will reconvene in the main lecture hall to summarize some of the key discussion points that emerged during the breakout sessions.
Please scroll down to select three sessions you would be interested in attending.
Breakout Sessions with Leaders in Health Research
Blake Richards – Machine Learning in Neuroscience
In this session, we will examine the use of machine learning for analyzing neuroscience data, particularly high-dimensional recordings of live neural activity. Neuroscience has seen an explosion of new techniques for recording electrical activity from the brains of awake people and animals. Theoretically, this data will allow us to finally understand distributed information processing in the healthy brain, and in-turn, dysfunction in neurological disease. It may also open the door to more advanced brain-machine interfaces. However, the high-dimensional nature of neuroscience data, and its lack of a priori structure, makes its analysis challenging. Machine learning is rapidly becoming key to meeting these challenges, with a variety of techniques being used to great advantage, including random forests, independent-component analysis, and deep neural networks. In this session we will review the latest techniques for recording live brain activity and discuss current machine learning approaches that are best suited to analyze them.
Jim Reilly – Machine Learning in Psychiatry
In this session we will discuss applications of machine learning to psychiatry and related fields. The ML paradigm is the new frontier in brain health research. For example, it is possible to predict the emergence of a coma patient by using machine learning techniques to detect the presence or absence of specific components in the EEG response to various stimuli. Machine learning analysis of the EEG has also proven useful in diagnosing and treating psychiatric illness, such as major depression and schizophrenia. This outcome is based on the fact these illnesses produce characteristic altered brain activity patterns that can be detected by machine learning analysis of the EEG. Further, the task of identifying an effective treatment for a specific individual who suffers from these illnesses has proven to be extremely difficult. Typically, an expensive and time-consuming trial-and-error approach is used to determine an effective treatment. However, by using ML analysis, it has proven possible to identify an effective treatment for depression at the outset of therapy. These examples will be discussed in further detail throughout the session.
Graham Taylor – Computational Peer Review
Millions of medical images are stored in the archives of hospitals and clinics, and are generally annotated with detailed patient-specific information (biopsy findings, treatment plans, follow-up reports). Being able to retrieve and describe images of similar cases (from past patients) when examining the images of a new patient can help physicians to make more informed decisions (more reliable diagnosis), decisions that are critical to both the patient’s well-being and to the costs of the healthcare system. The ability to correlate new patient images, with known (expertly diagnosed) cases of the past, can assist experts to avoid missing malignancies. Such a technology will exploit large-scale datasets and compute resources and constitutes a new generation of clinical procedures that can be coined “computational peer review”. This session will concentrate on the emergence and growth of digital pathology in recent years which challenges existing deep learning-based approaches due to very high-resolution images (e.g. 50k by 50k pixels and larger). However, we are also free to discuss methods for efficiently searching for similar examples of other kinds of unstructured data as well as methods to exploit meta-data (e.g. automatic annotation).
Babak Taati – Computer Vision in Chronic Illness
My research applies computer vision systems to solve healthcare challenges in rehabilitation and the management of chronic conditions. Several of my current projects include human pose tracking, human face/expression tracking, or human movement tracking in natural settings, such as in the home or in long-term care. Examples of my current projects include longitudinal vision-based monitoring of gait in long-term care homes to continually estimate the short-term risk of falling for each resident; vision based monitoring of facial expressions in older adults with moderate to severe dementia who cannot verbally express their pain; and vision-based monitoring of parkinsonism and levodopa-induced dyskinesia. It is difficult, time consuming, and expensive to collect data from patient populations. So, I am interested in transfer learning methods that could take advantage of pre-trained models trained with large amounts of data from healthy adults. I am also interested in problems related to dataset bias and covariate/label shift (e.g. from one hospital to another), and sensor fusion (temporal and non temporal data).
Laura Rosella – Machine Learning for Population Health
In this session we will discuss the application of risk prediction models for population health applications – that is to inform strategies that inform system level impacts on the health of populations and health systems (versus one patient at a time). To date, ML applications have been concentrated on healthcare and clinical decision-making. The application for these approaches for prevention of disease and the sustainability of health care systems – is limited. Most applications of ML neither capture the broader determinants of population health nor are they designed to address complex health system challenges (such as population-based prevention programs or targeted efforts at improving health equity). Here will work through some examples, and promising directions on how we can use ML techniques to improve health at the population level.
Charles Victor – Applications of Machine Learning and AI in Remote Monitoring Devices for Discharged Patients
Health care institutions are under increasing pressure to discharge patients early provided patient safety is not compromised. Benefits of early discharge following a surgical procedure are improved patient outcomes and patient quality of life due to recovery in a less stressful and more familiar environment, but also reduce hospitals’ per patient costs. The risks associated with early discharge such as negative or catastrophic outcome resulting from an inability to monitor the recovering patient need to be mitigated. Monitoring of home patient recovery has traditionally been through follow-up visits or phone calls. Modern proposed methods include patient vital sign telemetry, where continuous monitoring of key patient vital signs such as heart rate and blood pressure are recoded in real-time and sent to the discharge hospital for monitoring. However, these modern methods are still in early phases and being tested, challenges include “alarm fatigue” and information management (i.e., how does a team of clinicians monitor continuous vital sign feeds of hundreds of patients while continuing care on patients in hospital). Can machine learning and artificial intelligence be used to identify patients truly at risk of a negative post-surgical outcome and create algorithms embedded in remote monitoring devices that will alert the appropriate care provider. In this session we will present a randomized controlled trial currently under way in Ontario that is examining the impact of remote monitoring on cardiac post-surgical patient outcomes. We will discuss the potential for ML and AI techniques to utilize the data from this study to improve remote monitoring devices.
Cedric Manlhiot – Machine Learning for Cardiac Health and Hospital Management
At UHN, the Peter Munk Cardiac Centre (PMCC) performs more than 12,000 heart and blood vessel operations and 240,000 diagnostics and therapeutics procedures, and sees 163,000 out- patients in clinic per year. We have designed and built a PMCC Digital Cardiovascular Health Platform that will consolidate, clean, link and prepare data from all of the clinical systems (40 data sources) used in the care of these patients, in the most comprehensive way possible. We anticipate that all clinical notes, blood tests, imaging and genomic data that is generated while caring for these patients will be available for analysis, in real-time. The internal data in the PMCC data lake will be linkable to provincial and national data assets such as the Ontario Lab Information System (OLIS, all blood test results), Ontario drug benefit program (ODBP, tracks prescriptions) and administrative data from the Canadian Institute of Health Information (CIHI), thus providing a complete view of patients’ medical data. In partnership with the Vector Institute, we will use this data and ML approaches to improve the efficiency of hospital operations, personalize individual patient management and discover new associations that could lead to novel therapeutic options and improved outcomes for patients with life threatening heart and blood vessel diseases.
Anne Martel – Computer Vision for Precision Medicine
There is growing interest in the application of machine learning to the analysis of digital pathology images. Automated algorithms capable of finding small regions of cancer in whole slide images (WSIs) or quantifying the expression of specific antibodies in cells offer the potential to speed up pathologists workflow. Significant progress using convolutional neural networks has already been made in this area and in this session we will concentrate instead on the use of digital pathology to support precision medicine. This involves the use of image information to predict patient outcome and guide treatment decisions. For example, in one of our current projects, we analyzing the WSIs of surgically removed non-invasive cancers in order to predict which patients will go onto have an invasive breast cancer; those women with a higher likelihood of recurrence could be advised to undergo more aggressive therapy than those at lower risk. There are several unique challenges posed by this kind of task. A WSI is huge (a typical slide may be >10Gb) and relevant information spans multiple scales. Also, we only have labels at the patient level so we need to determine what part of an image is relevant to the prediction task. We will review various approaches that have been used to overcome these problems and will discuss new avenues to explore.