This article is the first in the ‘Technology, transformation and health care’ series created in partnership with AMS Healthcare. These solutions-focused articles will focus on emerging technologies and their potential for transformational change in our health-care system.
Since the release of OpenAI’s artificial intelligence tool, ChatGPT, last November, speculation over the possible applications of these types of technologies has run wild, with researchers exploring possible uses from writing the next great novel to detecting cancerous lesions.
ChatGPT is the latest variant of a language model known as GPT-3 that uses natural language processing (NLP), a subfield of AI, that enables computers to process human language in the form of text or voice data to “understand” its full meaning. The use of NLP in health care is an area of increasing inquiry. ChatGPT has been used to draft treatment recommendations for clinicians and patients with astonishing accuracy. One recent U.S. study has even used NLP to predict and diagnose early signs of dementia by analyzing patients’ speech patterns.
The most reliable application of NLP in health-care settings to date largely has been administrative, for example to organize electronic health records. But as these technologies get more and more sophisticated, the role they will play in health care likewise expands.
Take, for example, MuScRAT, a particularly successful AI tool that has been in use at St. Michael’s Multiple Sclerosis (MS) unit for the past five years and is now undergoing further development to expand its clinical abilities.
“Using Natural Language Processing, what MuScRAT does is it actually runs through clinical notes,” explains Muhammed Mamdani, vice president of Data Science and Advanced Analytics at Unity Health Toronto. “It can cycle through seven years of notes, for example, in about two or three seconds. And it actually will assemble all the pertinent information that the clinician needs into a visual graph.”
MS is a relapsing and remitting chronic central nervous system disorder that presents with a broad range of severity. Though there is no known cure, there are a number of treatment options that can improve the condition in some patients. The chronic and fluctuating nature of MS means that patients often have long and complex medical histories for clinicians to review.
“We have a huge MS clinic, bigger than most clinics around the world,” says Jiwon Oh, medical director of the Barlo Multiple Sclerosis Program at St. Michael’s Hospital, estimating that the clinic sees nearly 9,000 patients each year. “For neurologists who are caring for these chronic patients, some of whom have had this disease for seven or eight years, it’s really painstaking to go through all of the clinic notes to get an idea of what has happened with the patient.”
On top of translating important events in a patient’s medical history into an easy-to-read visual timeline, MuScRAT can read through decades worth of electronic records to render what is known as a Kurtzke Expanded Disability Status Score, or EDSS, a neurological disability score ranging from zero to 10. The first levels, 1.0 to 4.5, refer to patients with a high degree of ambulatory ability, 5.0 and onwards describe more severe stages of MS-related disability with 10 being death due to MS.
Oh says that although an EDSS score is an imperfect measurement, it is still the most common measure in clinical trials to gauge MS severity. It helps clinicians determine how stable a patient’s disease is and the course of treatment, as well as who may be eligible for certain therapies.
Without MuScRAT, a research coordinator would have to look through each clinical note to find either the clinician’s score or derive a score based on neurological exams.
“We’re just drowning in data and information.”
“The tool allowed us to quickly extract these disability scores from clinical charts, which is really helpful from an efficiency standpoint,” Oh says. “It’s helped clinical care in that sense because it allows people to get a quick snapshot of a patient’s entire disease history.”
The tool is particularly useful for residents and new hires trying to familiarize themselves with a patient’s disease history, she says. “You can provide much more informed and better care for your patient, rather than rummaging through seven years of notes on paper or even an [electronic medical record]. I think it’s the sort of thing that will help our clinicians provide better and more efficient care, because we’re just drowning in data and information.”
The clinic is working to use MuScRAT to expedite other administrative tasks as well.
“We have these amazing people called drug-access navigators in the clinic, who help clinicians to get approval for certain drugs for patients. To accurately populate an insurance form, you probably need at least 30 minutes of chart review for one patient. But with MuScRAT, we’re able to do it in five minutes,” she says.
Mamdani and his team see further potential for MuScRAT in the near future and are working on a new model that will be refined for deployment by the end of 2023.
“We’re in the process of integrating a risk-prediction model into MuScRAT,” Mamdani says. “Basically, what it does is it predicts disease relapse within the next two years.”
MuScRAT’s prognostic tool uses a combination of NLP, like the one it uses to derive an EDSS, in addition to other deep learning AI technologies. It takes all of a patient’s clinical and demographic data – age, sex, comorbidities, medications, how many relapses the patient has had, lesions from past MRIs and more recent scans – and uses that information to predict an MS patient’s risk of relapse.
“The model captures more than 70 per cent of all patients that will experience significant disability in the next two years and ‘gets it right’ more than 80 per cent of the time,” says Mamdani.
Oh says that there are more than 20 different treatments available for relapsing MS, some of which are new while others that have been around for decades. However, there are different side effects and potencies to each therapy. “Some of the side effects from these therapies can be pretty serious, even lethal,” Oh explains. “There’s a lot of treatment decision-making that clinicians make based on their clinical judgment of how severe a patient’s disease is. That’s where an accurate prognostic tool would help. If a patient clearly has negative prognostic factors, then a clinician wants to start them on the highest efficacy therapy.”
For now, MuScRAT is designed specifically for MS patients, but Oh says that AI tools that are able to prognosticate are becoming increasingly common. “It’s clear that you can use MuScRAT to quickly zero in on information, so I think there’s wide applicability in medicine in general.”
But how long will we have to wait until AI tools are helping doctors to make clinical calls? Oh says she is confident in the technology, but a lot of it will boil down to how willing physicians are to trust AI in the “hardcore, decision-making matters” and development of the technology.
“There’s this whole black box, unexplainable part of AI that doesn’t sit well,” Oh says. “I think clinicians deep down have a bit of distrust when it comes to computers making those decisions for us based on tons of data. That will be a bit of a hurdle in terms of acceptance from clinicians, even when AI algorithms seem to make decisions more accurately.”
The comments section is closed.