Innovative Projects Realized

Analysis of infusions pump usage data to reduce drug errors and enhance maintenance

An infusion pump is a medical device that automatically administers set amounts of drugs into a patient intravenously. Drug errors occur when an incorrect dose or dose rate is programmed. To avoid such errors, a drug library is created providing dose limits for every drug. “Smart” infusion pumps provide a tracking system, saving all pump usage data, messages and errors in a database. A study on “smart” infusion pumps indicates that drug errors and drug events were detected by the drug library; however, poor compliance to the library did not reduce drug error rates if users failed to adjust dose based on the drug library. Research will be conducted, leveraging the data provided by “smart” infusion pumps to improve compliance, and in turn reduce drug errors and associated complications. In addition, research will be conducted to use infusion pump data to enhance maintenance schedules and potentially predict device failures.

Développement d’un système d’acquisition de données sportives & gestion informatique de la clientèle pour les centres sportif

Le centre d’escalade Vertige développe présentement un système de gestion de centres sportifs. Ce système se voudrait être combiné avec une technologie sportive innovante servant à analyser les performances de grimpeurs et/ou de sauteurs dans plusieurs optiques. Pour le grimpeur ou le sauteur, cela servira à améliorer la rétroaction aux entrainements, à quantifier les séances pour l’optimisation de celles-ci et à avoir des statistiques sur ses performances. Pour le gestionnaire du centre, cela permettra de savoir quels modules sont les plus appréciés et ainsi d’adapter le centre à sa clientèle. Ça servira aussi à gérer les abonnements et entrées de façon efficace et permettra aux employés de se concentrer sur le service client plutôt que sur l’admission de ceux-ci et la gestion financière.

VR-based testing station for impairment screening

In this project, a VR-based testing station for impairment screening will be implemented. The station includes a Virtual Reality (VR) goggle (to be updated to Augmented Reality, AR, later), biophysiological measurement sensors, and an integration algorithm to integrate the result of measurement with scene construction of the VR system to implement dynamic scene rendering. The project will undergo several steps including basic scene construction and depth creation to model simple scenarios for the user and to research the implementation (and effect) of different related tests on level of impairment; sophisticated rendering algorithm creation to automatically design scenes based on the application; and dynamic scene construction to receive feedback from other measurement sensors and to implement a dynamic algorithm to update the scenes based on user’s reaction.

Understanding cell-cell interactions with deep learning-based profiling

The aim is to understand how fibroblasts, the most common connective tissue in animals, and cancer cells interact with each other through image analysis. These co-culture imaging screens, containing fibroblasts and cancer cells, will help identify novel signaling mechanism involved in cancer. The objective is to apply deep learning techniques to these image-based assays to study interactions between and identify novel therapeutics that can make cancer therapies more effective.

Question-to-question semantic similarity for Question Answering System

Question Answering (QA) system automatically answer questions raised by users in natural languages, and it is a crucial component of a human-machine conversation system. A typical QA system collects human written question-answer groups and structures them in a database system. However, in order to answer questions that are semantically similar to the questions stored in the database but are worded differently, the QA system needs to be able to calculate the semantic similarity between different questions. In this research project, the intern will explore different techniques used in question-to-question semantic similarity measurement and try to improve upon the state- of-the-art performance. From participating in this project, RSVP Technology Inc. could seed for more opportunities to collaborate with Canadian community to improve the quality of QA systems used in many other fields and products, such as customer service chatbots and smart home device. Further, this project could serve as the foundation for next step research and development.

Off-Policy Reinforcement Learning (RL) for a Production Robotics Application

Kindred offers eCommerce retailers a solution to assist with rapid order fulfilment from their distribution centres. The solution (SORT) is a combination of a so-called put-wall and a humanoid robot. The robot picks up items from orders, scans them, and puts each item in a cubby of the put-wall according to the scan code. The robot comprises a gripper, a 6-degree-of-freedom arm, and a stereo vision module, as well as other electronics and mechanical housing. The proposed research will explore machine learning techniques based on reinforcement learning to feed data recorded from Kindred’s production robots back into learning algorithms in order to generate new better ways for those robots to pick, scan, and stow those eCommerce customers’ orders.

Deep Learning/Computer Vision for Robotic Manipulation

Research is rapidly progressing in enhancing the Artificial Intelligence of Robotics. One backbone of this rapid change lies in Deep Learning. Deep Learning refers to new algorithms that are capable of learning behaviors after being trained by several thousands or even millions of examples of what should be done given an input. My project will be doing computer vision related research in this field. Computer vision mainly refers to tasks where a significant part of input is from a visual source, like a camera. For example, a robot would receive a video feed from a camera as input, akin to how a human would see from their eyes. My research is to delve into new ways to train a robot such that it can perform human like tasks (such as picking up objects or inserting pegs in holes) primarily using video feed input from a camera. TO BE CONT’D

Audience Allocation to Retail Geo-clusters

Based on the user’s geo-location, timestamp and other attributes (eg. time of day, past visit history and app behavior categories, etc.), a machine learning algorithm can be developed to find which cluster the users belong to. Overall, the data of geo-location and timestamp are used to roughly locate the potential clusters. This project will involve some techniques and algorithms like cloud computing i.e Google Cloud Dataproc, sliding windows, histogram and machine learning algorithms. The challenge of first phase would be coming up with a good way of estimating the number of clusters. Then by applying all the above techniques, the decisive attributes can be decided and combined to determine which cluster the users belong to.

Real-time object recognition on wearable devices

The goal of the project is to implement real-time state of the art object recognition models on wearable devices. These devices aim to help people living with a visual disability by providing a description of their outdoor environment and offer navigation guidance. This would improve the experience of the users by allowing them to perform usual day-to-day tasks with much more ease and safety.

Assessing and Addressing Health Disparities Related to Utilization of Preventive Care Services in Ontario

Health disparities arise as a result of long-standing societal disadvantage and discrimination. As machine learning models become more popular in the healthcare sector, understanding of current health disparities becomes even more critical. Without careful management of existing biases, the models can inherit and amplify health disparities, leading to highly undesirable clinical outcomes. This project focuses on health disparities in access to preventive care services. Preventive care services such as screening and preventive medicine allows for early diagnosis and timely interventions. This project aims to provide an understanding of if and how patterns of preventive care utilization aggravates health disparities in Ontario, by employing advanced data exploration and visualization techniques. After establishing such a relationship, this project also provides an individual risk profiling tool to assess the efficacy of preventive services, using advanced feature representation and deep learning techniques.

Electrical Load Forecasting

Load forecasting is an essential activity for a company like Hydro-Québec. It is necessary for objectives as varied as the management of production or the management and maintenance of the electricity network. Any significant forecasting error can result in reliability issues, loss of opportunity, or additional costs to the business. On the other hand, a good prediction would allow Hydro-Québec to generate additional sales in neighboring markets. With the deployment of its Advanced Measurement Infrastructure (AMI), Hydro-Québec now has a significant amount of new consumption data. This data can be used to improve demand forecasting, increasing reliability, decreasing expenses, and potentially generating new revenue.TO BE CONT’D

Role of transthoracic impedance and current in synchronized electricalcardioversion

Synchronized cardioversion is a medical treatment that applied an electrical pulse to restore a normal heart rhythm is patients with an abnormally fast heart rate or cardiac arrhythmia. A successful cardioversion is dependent on the amount of electrical current that reaches the heart, which depends on the strength of the electrical pulse and the transthoracic impedance (electrical impedance of the body). If a cardioversion is not successful, additional attempts

are often made; however, repeated delivery of a large electrical pulse is not desired. Increasing the strength of the electrical pulse is also not ideal as it increases the chances of complications. This research will investigate the role of transthoracic impedance on cardioversion, which will also include the effect of different paddle placement. Outcomes of this research will help improve the efficacy of cardioversion, increase the success of treatments, while minimizing the strength of the electrical pulse.