Development of cardiac specific machine learning infrastructure
Analytics for Life, Inc. (A4L) is an early stage medical device company that specializes in the development of technologies to analyze patient physiological signals in order to evaluate cardiac performance, status and risk. A4L’s core competencies include identifying and developing mathematical features from physiological signals and assembling these features into clinically informative formulae using machine learning techniques. A4L has used third party machine learning tools (open source and licensed products) for the formula generation aspect of the product development cycle.
Specifically, A4L has used these tools to demonstrate the feasibility of computing left ventricular ejection fraction, cardiac ischemic burden and other cardiac performance/status parameters for simple to collect, non-invasive physiological signals (surface voltage gradients, SPO2, Impedance etc.). As a result of this experience, A4L have learned the benefits and insufficiencies of these tools for A4L’s specific purposes. A4L plans to file with the U.S. Food and Drug Administration (FDA) an application for approval of a physiological signal collection device and will soon afterwards be seeking market clearance for products assessing cardiac health emanating from the machine learning process. A4L believes it can build a machine learning tool specifically tailored for cardiac evaluation based on experience with the tools used to date.
This A4L-specific machine learning paradigm will search only relevant mathematical spaces, cutting down on time and CPU power needed to iterate to solutions and will allow for an assessment of a much wider array of potential solutions. Furthermore, this A4L-specific machine learning paradigm will provide a controlled and validated system that can be audited and evaluated by regulatory bodies, something that is not possible with the current machine learning tool(s). A4L proposes a hybridization of paradigms within a set mathematical space. This will create efficiency in the search, and therefore more searches can be performed in the same period of time. This will lead to more solutions being available for evaluation, resulting in more accurate and efficiently produced end solutions. If successful, this new paradigm will allow for simple, non-invasive, rapid and relatively inexpensive cardiac diagnostic capabilities, bringing tertiary care diagnostics to primary care settings and disrupting the current infrastructure and capital cost-centric model of diagnostic delivery.
Development of severity index of exacerbation for COVID-19 symptoms from abnormal respiratory patterns
Struggling to breathe has been a challenge that existing monitoring devices have faced, for identifying early symptoms of respiratory diseases such as novel Coronavirus (COVID-19). Measuring respiratory rate alone would result in potentially missing if a patient was unable to inhale a full breath.
To overcome this challenge, a novel technology solution invented in Canada is used through an Intelligent Bed Sheet, with the ability to continuously monitor if a patient has normal, shallow, or irregular breathing.
This collaboration with York University will revolve around developing a Severity Index by identifying earlier when a patient is struggling to breathe. Using the most advanced electronic fabric technology, the SOSCIP Platform will support computing challenges by using large amounts of breathing data in a constantly changing environment, to precisely identify before a patient’s breathing becomes more severe.
Distributed and scalable search in enterprise databases
Google search, and other search engines such as Bing and Yahoo!, provide a convenient way to find Webpages that contain various keywords or are related to particular topics. For the purposes of searching, Webpages are essentially loosely structured paragraphs of text. However, much of the world’s high-quality enterprise data are structured into well defined tables containing sets of well-defined columns.
One consequence of structured database design is that information about a single entity may be scattered across many columns in many tables, and must be stitched together in a meaningful way when answering user queries. This turns out to be significantly more difficult than finding Webpages or text documents containing various keywords.
As Dr. Surajit Chadhuri (a Distinguished Scientist at Microsoft Research) recently argued in a keynote talk at the IEEE Data Engineering conference, search over structured databases has fallen behind search over unstructured data. In the proposed research, we will develop a powerful and intuitive search system, akin to Web keyword search, for structured enterprise data. Our system will empower nontechnical users to explore enterprise databases and turn big data into actionable insight, just as Google search has empowered society to explore the Web.
Distributed Deep Learning and Graph Analytics Using IBM Spectrum Computing Solutions
Deep learning is a popular machine learning technique and has been applied to many real-world problems, ranging from computer vision to natural language processing. In most cases deep learning outperformed previous work. However, training a deep neural network is very time-consuming, especially on big data. A popular solution is to distribute and parallel the training process across multiple machines. Indeed, the race is on to parallelize deep learning! Industry and academic research teams around the world are trying to make deep neural networks train as fast as possible on farms of GPU capable servers. We are working with our IBM partners to help develop advanced scheduling and messaging techniques for distributed deep learning. In addition, we will develop two real-world applications of distributed deep learning to demonstrate the efficiency and effectiveness of distributed deep learning. In one application, we address the video surveillance problem of tracking a moving target over a network of video cameras with partial or no overlaps in their coverage. We will use a deep learning approach to identify multiple pedestrians in each video frame, and a particle filter to track moving pedestrians. In the second application, we address the problem of fraud/intrusion detection. We will use graph-based detection that considers relationships between objects or individuals. Graph-based approaches are powerful because they do not operate on objects or individuals in isolation, but also consider their network information. We will emphasize on graph-based fraud detection methods that have a number of applications and potentially large impacts.
Dynamic microscopy image processing and analysis for infectious diseases, diagnosis and treatments
In vaccine and therapy development for infectious diseases, advanced optical imaging is used to measure the interaction of virus and the host cell. Current microscopes are relatively slow that will only provide a “snapshot” of the biological interactions. In order to develop diagnostic methods and effective treatment, it is necessary to image these dynamic interactions continuously like a movie. Additionally, the stream of images will also need to be processed and analyzed rapidly using new computation methods. To address such a challenge, we plan to develop high-speed microscopic imaging instruments and related image processing and analysis technology. These technologies will enable us to build a customized microscope capable of high-speed quantitative imaging of virus-host interactions in live cells for infectious disease diagnosis and therapeutic treatment research. This project module will primarily focus on image processing and analysis algorithm development.
Efficient deep learning for real-time traffic event detection
Miovision is interested in designing the first affordable, low-power, energy efficient real time traffic event detection system that can be installed without the need to be powered by the grid nor the need to be connected directly to city installed infrastructure. Deep learning for traffic event detection can provide overwhelmingly superior accuracy and addresses most of the real-world scenarios that make competing detectors unsuitable for customer adoption. The challenge with deep learning is its complexity, which is currently infeasible for a self-powered real-world embedded detection system. Working with Dr. Alexander Wong and the Vision and Image Processing Lab at the University of Waterloo, the goal of this project is to develop technologies that can significantly reduce the complexity of deep learning for traffic event detection, while maintaining its accuracy and market fit, so that it can be deployed on a low-cost and low-powered hardware platform.
Electrochemical Fischer-Tropsch synthesis of renewable liquid fuels from CO2
Renewable electricity costs have been rapidly declining, enabling clean consumption of energy in many sectors. However, there is still demand for energy-dense liquid fuels, such as in heavy freight and air transportation. In this project, we will harness machine learning to develop technologies that enable the synthesis of liquid fuels from carbon dioxide and/or synthesis gas using renewable electricity. Industrially, liquid fuels can be synthesized from a mixture of carbon monoxide and hydrogen called synthesis gas (syngas). However, this process requires high temperatures and pressures, and is itself responsible for significant greenhouse gas emissions. We propose the use of electrocatalysis to produce these liquid fuels. To accomplish this, we will use computational modeling and machine learning methods to design electrocatalysts that efficiently convert CO2 or syngas into dense chemical fuels. These computational efforts will be validated through a parallel experimental approach that includes the fabrication of new catalyst formulations and the construction of prototype electrochemical flow cells. This project will enable the synthesis of clean, energy-dense liquid fuels that can replace the use of fossil-derived fuels in industry and transportation sectors. This project is a logical extension of the existing CO2 related projects previously underway with SOSCIP. This project will allow the research to achieve the next milestones in the overall goal to achieve beneficial conversion of CO2.
Fast and accurate biophotonic simulations for personalized photodynamic cancer therapy treatment planning
There are many medical uses of light, both for diagnostic purposes (medical imaging) and for treatment. We will build a very fast and accurate simulator of where light inserted with fiber optic probes travels within a person’s body; by using this simulator we can enable a promising new cancer therapy, among other medical applications. The key use of our light simulator in this project will be for photodynamic therapy (PDT), a promising new cancer therapy. It uses non-toxic light activated drugs (called a photosensitizer) which is harmless until it is activated by light of a certain wavelength; where the activated photosensitizer destroys cells. Hence if we can localize the light to a cancerous tumour, we can destroy it with minimal damage to surrounding healthy tissue and with fewer side effects and less cost than would be achieved with ionizing radiation therapy or surgery. PDT is used today to destroy “superficial” cancers, such as those on the skin where it is easy to localize the light to the tumour.
PDT can also be used to destroy tumours within the body by inserting fiber optic probes through needles into the body. The key challenge is that the light reflects, refracts and is absorbed in complex ways when it leaves the fiber, making it hard for a physician to determine where to place the fibers and whether the tumour will be completely destroyed. We aim to fill this gap by developing a fast and accurate simulator that can determine the light density throughout a person’s tissue that will result from a given fiber optic probe placement and light input intensity. To do so we will compute the path taken by hundreds of millions of simulated photons — a task so computationally intense that we will use an unconventional IBM computing platform that uses special programmable hardware to offload key calculations from the conventional processor. We believe that by doing so we can speed up the computation and reduce the power required by a factor of 60. This large speed up will allow us not just to simulate one probe placement but many possible light probe placements — we seek to return the best probe placement and the expected treatment results to the physician in less than an hour. Overall this project will pave the way for a much-needed new, less invasive, more effective, and lower-cost cancer treatment. It will also use a new style of computation where special hardware does much of the computation, instead of a general purpose CPU, showing how such an approach can be used to make future computers faster and less power-hungry.
Forecasting Covid-19 Epidemic in Canada with Spatial-Temporal Models That Exploit Population Behaviour on Twitter
The Covid-19 pandemic is creating unprecedented damage to the public health and world economy. Being able to accurately forecast the spread of Covid-19 is critical for the federal and provincial governments of Canada to devise policies and measures maximally protecting the lives of Canadians and rapidly reviving the Canadian economy. Not only important at present, developing advanced epidemic projection algorithms and techniques also helps prevent future epidemics of other infectious diseases. In this project, we set out to develop advanced forecasting algorithms for the spread of Covid-19 across Canada. Specifically, we will exploit the behaviour information of the population revealed on social media as well as the correlation of the Covid-19 spread across different regions of Canada. Our development will integrate modern AI techniques in data analytics, machine learning and natural language processing with the conventional mathematical models for infectious diseases. The developed algorithms will be hosted on a web platform, which will provide accurate daily predictions of Covid-19 spread and release them to the public. These predictions will assist the governments to strategically adjust policies for the protection of Canadian lives and the revival of the Canadian economy. Individuals, families, schools and businesses will all benefit from this new source of information in their planning processes.
Full Monte: fast hardware for Monte Carlo biophotonic simluations
There are many medical applications of light, such as bioluminescent imaging (BLI) and photodynamic therapy for cancer treatment (PDT). These applications require rapid and accurate simulation of how light will scatter and be absorbed in complex human tissue in order to image tissue (e.g. BLI), or to determine where a light-sensitive drug will be activated and destroy cells (e.g. PDT). This project will use agile computing (FPGA hardware) to simulate light propagation very accurately with the Monte Carlo method but much more quickly and power-efficiently than in a conventional computer. The first aspect of the project is to complete our prototype agile implementation of this simulator and scale up its performance and the problem sizes it can handle. The second aspect is to complete a workflow using this simulator along with meshing, visualization and light source placement optimization to make a clinically useful PDT treatment planning system.
Generalized heterogeneous radio signal processing
A new generation of radio telescopes is opening new windows on the Universe, allowing astronomers to observe the cosmos in unprecedented ways. Powered by the ongoing revolution in computing, these new telescopes operate at the cutting edge of digital technologies. New algorithms are being developed at a spectacular pace, and we are forging a new partnership with the Markham branch of Advanced Micro Devices (AMD) to add to these, developing new tools and opening new possibilities in radio astronomy, software defined radio, and similar telecommunications technologies. High-cadence mitigation of Radio Frequency Interference (RFI), advanced Digital Beamforming techniques, and Dynamic Spectral reshaping will all be developed and ported to an open software framework, allowing it to be used on a wide variety of computational and signal processing hardware.
Generative models for autonomous, video-based DJ systems
RaveDJ is all about enabling people to create music with AI. Over the past few years our team has worked tirelessly to make RaveDJ the world’s first fully autonomous DJ system. A system that can seamlessly mix together a set of songs to produce the optimal mix to listen to as well as unique combinations or mashups of two songs into one. RaveDJ is currently producing around 10-50 TB per day of mashups and mixes for people all around the world. RaveDJ is hosted on our website www.rave.dj as well as our app Rave (get.rave.io/download) which has millions of users from around the world. SOSCIP’s platform will enable us to investigate ways of using bleeding edge machine learning techniques to navigate around existing inefficiencies in our existing system, as well as to take a leap beyond its current level of performance with new research advances. SOSCIP’s platform is exceptional for helping us achieve these goals as it can support training of deep neural networks on very large datasets of music videos. This is a requirement for analyzing and generating thousands of high-resolution videos at record speed. The platform that SOSCIP provides is perfect for running the experiments that are required to produce models that can perform these tasks efficiently as well as addressing new and novel problems in using AI for creative applications. The collaboration is between Rave, Prof. Graham Taylor and Prof. Stefan Kramer at University of Guelph.
Genetic variation and structure-based drug polypharmacology: multiscale structural pharmacogenomics
Over the past 25 years, drug discovery efforts have been aimed at identifying small organic molecules that exert a strong effect on a single protein that corresponds to the biological target of a disease. Although designed to fulfill a single biological function, small molecule drugs can incidentally interact with hundreds of the ~20,000 known human gene products. These off-target interactions may lead to side effects, new therapeutic effects, or otherwise affect drug response. Mapping a drug to all of its human protein interaction partners is known as polypharmacology. Thus, it is advantageous to understand the polypharmacology of each drug in development prior to clinical testing in order to better anticipate off-target effects that could potentially cause toxicity. However, to address drug polypharmacology experimentally is prohibitively expensive in time and monetary cost. Predicting a drug’s polypharmacology is Cyclica’s central mandate. Our drug discovery platform is based on proteome-wide screening, and centered around PROBEx, a cloud-based software solution that simulates a drug’s interaction with 100,000’s of proteins on the basis of their molecular structures. PROBE x maps out the possible off-target protein partners and consequently pharmacology for any pharmaceutical or nutraceutical compounds, to service academic research institutes, hospitals, and pharmaceutical companies developing new disease therapies. Pharmacogenomics is the study of how genes affect a person’s response to drugs by combining pharmacology (the science of drugs), and genomics (the study of genes and their functions) to develop effective, safe medications that are tailored to a person’s genetic makeup.
This field, commonly termed “personalized medicine” has recently emerged and gained significant attention from many companies, including IBM Watson Life Sciences. To our knowledge however, information pertaining to the structural proteome and its relation to drug binding, has not yet been systematically leveraged to improve pharmacogenetic analyses. This technology gap represents a unique opportunity for Cyclica to contribute to personalized medicine, by working with SOSCIP and the IBM cloud analytics platform. Cyclica can integrate its drug polypharmacology predictions with clinical and genomic data to build models that can explain and predict variation in drug response between different individuals. Through Industrial Partner: Cyclica Inc. this program, we will create new tools that match patients to the pharmaceuticals best suited for their distinct genetic features. Ultimately, the effort will improve outcomes and limit the loss of valuable time associated with trial-and-error medicine.
Getting ahead of the curve: a novel way to find people who are likely to be asymptomatic carriers of COVID-19 before they infect others
The spread of COVID-19 has been slowed by physical distancing, self-isolation, lockdowns, masks and travel restrictions. Doing more testing and contact tracing of known cases has helped contain the spread. However, these approaches are ‘behind the virus’ – by at least 5-10 days. We need a way of getting ahead of the virus, before it spreads further. We know that about half the infections are transmitted by people who do not have any symptoms, and this is sufficient to cause a growing epidemic. Therefore, we need to find, isolate and test at least some asymptomatic people. Like intelligence officers hunting hidden terrorists, we need to prevent attacks before they happen, instead of chasing the culprits after the fact. In other words, we need to identify carriers before outbreaks occur, not after. Our team proposes a new way to do this.
The overall goal of this project is to develop models and an intervention prototype to predict which individuals are most likely to be exposed to COVID-19 and are therefore most at risk of onward transmission. We will apply statistical modeling and Artificial Intelligence (AI) methods to demographic, occupational, social networking and geolocation data. In simulations, we will test different ‘smart isolation and testing strategies’ that could be used by public health officials to determine which would most effectively reduce viral transmission. Applying AI-informed strategies could enable partial relaxation of confinement rules for lower-risk segments, and allow greater reopening of economic and social life without risking a second wave and further lockdowns.
Harnessing the diversity of phage displace libraries to capitalize on single domain antibodies with high target affinity and improved protease stability
AbCelex Technologies is developing a line of novel antibody-based products delivered as feed additives to poultry for the prevention of foodborne illnesses caused by Campylobacter and Salmonella. AbCelex’s single domain antibody (sdAb) platform technology is based on camelid antibodies, which have significantly higher thermal and protease stability, accessibility to target due to their small size (1/10 of conventional antibody) while providing affinities that exceed that of conventional antibodies. The current platform utilizes bioinformatics approaches that take into account genetic and protein diversity of the pathogens and in silico antibody engineering predictions to inform optimal design of sdAbs with competitive affinity, effectivity and cost as feed additives. In this project, we aim to combine the benefits of the natural scaffold of the camelid antibody and the expertise we have developed on understanding the interactions of these sdAbs with their target to develop computational libraries that we can then validate in vitro, ex vivo and ultimately in vivo.
For this purpose, we propose to collaborate with Dr. Mauricio Terebiznik, with whom we have been developing a detailed database of the target-antibody interaction mechanisms. He is an expert in cellular biology mechanisms altered by the pathogens such as Salmonella. The major barrier to achieve this aim is access to high performance computing and the SOSCIP platform is ideal to overcome this barrier. Furthermore, the postdoctoral funding will allow the candidate to build on the backbone of camelid antibodies rationally selected peptides that would provide target specificity and cross-reactivity across different strains of the same bacterial pathogen through this proposed collaboration.
High fidelity simulations and low-order aero-acoustic modeling of engine test cells
The testing and certification of gas turbines demand the well-controlled environment provided by an engine test cell. The resonant acoustic coupling between the flow generated noise from the gas turbine exhaust and the engine test cell impacts the quality and reliability of the engine testing and certification. Predictive modeling of flow generated noise using high-fidelity numerical simulations is central to an a priori acoustic assessment and for the development of noise-mitigating designs. As part of this effort, the Multi-Physics Interaction Lab and University of Waterloo will numerically study, with the help of high fidelity, large-eddy simulations of the SOSCIP high-performance computers, the acoustic noise generation in partially confined jets undergoing a re-acceleration through the test cell ejector system. As a direct outcome, the researchers will develop a low order Aero-acoustic model that will be used by our industrial partner to predict resonant acoustic models to within +/- 20% the frequency and amplitude of the coupling phenomena. This OCE and NSERC-funding project will permit the training and mentoring of four HQP for careers in science and technology within Canada.
High performance computing for assessing and mitigating the effect of extreme wind on building and cities
As the second largest country in the world, Canada’s diverse geography and climate increases our cities exposure to different types of natural hazards, such as snow storms, hurricanes, tornadoes and floods. The insurance industry estimates that insured catastrophic losses in North America average $80B per year. In Toronto (2005), for example. a single tornado event resulted in $500M loss. This is further compounded by changes in climate, population growth and aging infrastructure.
To maintain the prosperity of our communities, it is imperative that a comprehensive framework be developed to assess and mitigate the impacts of extreme climate on cities. The current project aims to develop a multi-scale climate responsive design framework that accounts for the complex interaction between buildings and wind (including hurricane and tornado). This computational framework, at neighborhood scale, models urban micro-climate necessary to assess the impact of changing city topology on the pedestrian level wind, air quality and to generate boundary conditions for small-scale simulations. At building scale, it develops a full numerical aeroealstic model (e.g. building model that flex) immersed in turbulent city flows, for the first time. This frame work when integrated with artificial intelligence based optimization procedures, allow optimizing tall building aerodynamics (shape) and dynamics (structural systems) appropriate for current era of booming tall building construction.
As a result, Ontario will save materials and energy in one of the most resource intensive sector, while enhancing the safety of Ontarians during extreme climate. For successful implementation of the framework, a high performance computing environment and experimental validations are necessary, which will be enabled by two unique research facilities in Ontario, Blue Gene Q and WindEEE Dome, respectively.
High-fidelity aerodynamic analysis of unmanned multirotor vehicles
The objective of the proposed research is to gain a better understanding of the complex aerodynamics of small multirotor vehicles, such as quadcopters. This will enable Aeryon Labs, a Canadian company and world leader in small unmanned systems, to improve the development cycles of their products and respond faster to customer needs. Multirotor vehicles are popular platforms for many remote sensing applications because of the relative ease to control them at hover. During fast flight, however, control becomes challenging, because of highly nonlinear aerodynamics. These nonlinearities are due to the small-scale aerodynamics typical for these vehicles, and the interaction of the flow fields of several rotors that operate in close proximity. The proposed research builds on existing research on multirotor-vehicle aerodynamics. In order to expand our understanding of the complex aerodynamics of multirotor vehicles, we propose to model a quadcopter using Computational Fluid Dynamics (CFD). The CFD results will be compared with the lower-fidelity predictions and experimental results. The investigation of the predictive method will benefit from experiments performed in the large low-speed wind tunnel at Ryerson University, flight tests and an existing collaboration with Aeryon Labs. As part of the proposed research project one postdoctoral fellow, one doctoral student and two MASc students will receive training in the area of applied aerodynamics. The results will benefit Aeryon Labs with superior design tools that will improve their product line. Small aerial systems represent a rapidly expanding market segment worldwide, in which Canadian companies, such as Aeryon, play an important role.