CANT-WAIT uses AI to reduce MRI wait times
January 28, 2021
VANCOUVER – The University of British Columbia Cloud Innovation Centre (UBC CIC) in November launched its CANT-WAIT project, which uses machine learning and natural language processing to better manage requisitions for MRI exams.
The tool was developed in partnership with the Lower Mainland Medical Imaging Program in British Columbia with the goal of predicting the priority of the patient and to identify any indications for contrast-enhanced imaging. The vision is to use the tool to match MRI requisitions with regional waitlist information, in an automated way, and to direct patients to sites with the shortest waitlist for the indicated priority, resulting in reduced wait times.
The AI system has been designed to help transform paper-based requisition workflows and to support the automation of a centralized requisition distribution model.
In the Lower Mainland of BC, non-urgent outpatient MRI requisitions are received at a centralized distribution office. Traditionally, ordering providers would send imaging requisitions to a site of their own choosing. This unorganized approach can lead to a concentration of requisitions at some sites, causing disparate wait times.
Patients across British Columbia and in other Canadian provinces can wait weeks or months for an MRI scan. The CANT-WAIT technology is designed to support getting patients to their exams in the shortest time possible.
“It’s an ongoing battle to meet wait list targets when requisitions don’t receive a priority until after they are received at a site,” said Dr. William Parker, radiology resident at the University of British Columbia and founder of SapienML.com, a company specializing in creating solutions using Big Data. “It’s a massive task, but with Canadian Association of Radiologist (CAR) guidelines, and a large volume of requisitions, this project can have a transformative and scalable impact on the triaging process, saving time, resources and money.”
As an answer to the problem, the UBC CIC, in partnership with SapienML, radiologists at Vancouver General Hospital and St. Paul’s Hospital, Vancouver Coastal Health and Provincial Health Services Authority devised an AI system that can predict a requisition’s priority and whether there is an indication for contrast.
Using machine learning, they can process the massive amount of incoming requisitions and analyze the information that they contain, leading to an output that could be used in an algorithm to enhance requisition management in a single or multi-site health region.
The solution, powered by Amazon Web Services (AWS), was built by students at the UBC CIC, which officially opened at the beginning of 2020. The only centre of its kind in Canada, and one of just 12 around the world, UBC’s CIC is connecting top university researchers, companies and experts at AWS to create new solutions to difficult problems that impact health and wellbeing.
To stream patients needing MRIs to centres with the best availability, Dr. Parker said that SapienML and its partners devised an algorithm that first figures out the priority level for each patient. The CAR Priority classification system is used in BC to triage Medical Imaging requisitions.
It is assisted in this work by natural language processing, as ordering providers will often jot in notes as free text in the requisition. The system is able to interpret the free text and differentiate between acute and chronic injuries.
For example, a patient in pain whose knee popped a day ago and can’t walk needs an exam sooner than one who has experienced chronic knee pain for many months. “We need to decide, based on probabilities, that one patient will probably need immediate surgery and the other one can probably wait a little longer,” said Dr. Parker.
The algorithm accounts for many details written on the requisition by the ordering provider, including age, whether the patient is pregnant, if it’s a follow-up visit, whether the patient has claustrophobia, allergies, and other factors.
SapienML and its team will look to refine the model further, continually improving it. “Once it’s perfected, we’ll roll it out to other places too,” said Dr. Parker.
Not only can it be used for MRI exams, but Dr. Parker believes that CANT-WAIT could be modified for use with other modalities, such as CT and ultrasound. These modalities also have published prioritization guidelines that could be applied by the tool. It might also have the potential to reduce waitlists for surgical patients, another bottleneck in British Columbia and across Canada.
To create CANT-WAIT, the UBC CIC used natural language processing services from AWS, Amazon Comprehend – which uses machine learning to uncover the insights and relationships in the unstructured data – and Amazon Comprehend Medical, to extract complex medical information from unstructured text.
SapienML made use of its own data extraction software, called, DataRig and SapienSecure to extract all of the MRI requisitions efficiently and to remove the DICOM data, leaving only the requisition to be transcribed by the clerks. The software was also used to evenly distribute the requisitions by body parts and scan types.
The actual process of building the algorithm took about three months. Well before this stage, however, AWS and the UBC CIC brought together all the stakeholders in the project, to gain agreement on the project and begin the flow of ideas.
“The whole process took about six months,” said Coral Kennett, digital innovation lead at AWS. She noted that AWS has a methodology for helping teams produce solutions quickly. “The idea is to move fast, in an iterative way,” she said.
Vancouver Coastal Health and SapienML are involved in a second project at the UBC CIC that is also showing great promise. Called L3-Net, it’s an AI-based model that analyzes CT scans of the lung to determine the presence of COVID-19 disease and to assess how severe a case the patient has.
The SapienML team (which is also led by Dr. Savvas Nicolaou and engineer Brian Lee) observed that COVID-19 is a respiratory disease that has a huge impact on the lungs of patients, and that the damage is often long-term.
The AI analysis is a potential adjunct to pulmonary function testing, as not all geographical regions – especially remote areas – have ready access to all tests. “But CT scanners are quite ubiquitous, and you can send the exams to radiologists using telemedicine,” said Dr. Parker.
Patients who survive COVID-19 but experience damage to their lungs often consult afterwards with respirologists. As another facet to the project, the UBC CIC is now working with respirologists at St. Paul’s Hospital, in Vancouver, to correlate COVID-19 findings with various forms of lung damage and deficiencies.
Dr. Parker explained that pulmonary tests involve assessing a patient’s day-to-day function, lung capacity and other tests, as well as tests of the efficiency of the lung at taking in oxygen and releasing carbon dioxide and other molecules from the body.
Even though these tests are needed and are the gold standard, the exams are often unpleasant for patients. “They can be an exhausting, annoying experience,” commented Dr. Parker.
On the other hand, CT scans are simple to administer, and could provide similar info. “The patient just lies on a table,” said Dr. Parker. “It’s easier.”
Despite the appearance of vaccines to combat COVID-19, Dr. Parker said it will take some time to eradicate the virus. “COVID-19 isn’t going anywhere,” he said. “It will be here our whole lives.”
He noted, “We will also be helping patients dealing with the after-effects of the virus, such as reduced lung function.”
For its part, L3-NET was devised in Vancouver using over 3,000 CT studies from patients around the world. It was a large project, involving 50 or more people at the university, from AWS, and from the Vancouver General Hospital, Vancouver Imaging (a large group of radiologists), Xtract.ai, Element AI, and md.ai.
The first step involved collecting the data – namely, the CT scans. The next stage revolved around labelling the sets, which is an intensive exercise involving skilled personnel. “We worked with radiologists from Vancouver General Hospital and the UBC medical students,” said Kennett. “Labelling takes a high degree of medical expertise.”
The annotations are broken into three categories:
- Lung segmentation
- Opacity segmentation and classification
- COVID-19 classification (such as probable, indeterminate, non-COVID-19)
The project is now in its second version, Dr. Parker said, as improvements were made to the first. The code has been made freely available on Github, and researchers and clinicians can even upload CT scans for analysis on the UBC CIC website.