Article In Brief
Artificial intelligence and machine learning technologies—which can use algorithms to look for relevant patterns in large, complex datasets—already are showing promise for pinpointing novel biomarkers for multiple sclerosis.
Multiple sclerosis (MS) could be ideally situated to benefit from artificial intelligence (AI) as a fairly large window of opportunity exists to use AI/machine learning (ML) technology to identify subtle but potentially significant signs and symptoms and perhaps flag patients who could benefit from early intervention.
”MS is a disease in which small disturbances in one or multiple body systems can start happening early, and unless you are paying attention to them, they may not by themselves be sufficiently worrisome,” said Sergio Baranzini, PhD, an AI researcher and professor of neurology at University of California, San Francisco. “We know that if we are successful in identifying patients earlier and can start treating them earlier, that can improve the outcomes later on, and that is not true for some other conditions.”
AI/ML technologies, which in varying forms involve algorithms designed to look for relevant patterns in large, complex datasets, already show promise for pinpointing novel biomarkers for MS throughout the body, including in the brain, eyes, and gut.
While there have been advances in the treatment of MS in recent years, much needs to be uncovered about the underlying biomechanisms, genetics, and risk factors. Diagnosis of MS often is delayed because the disease presents in multiple and subtle ways, and even with a firm diagnosis, it is hard to predict its future progression or degree of disability.
MS researchers using AI/ML technology hope their work will lead to the development of clinical tools and protocols that will not only enable early diagnosis and prognostication but also promote more individually tailored treatment.
Dr. Baranzini said AI/ML, when implemented using established databases, has the potential to connect the dots gathered over years in a patient’s electronic health records. For instance, a patient may mention in passing to their primary care doctor that they are experiencing fatigue or an occasional bowel disturbance, which can seem like ordinary complaints but might also be part of the prodromal stage of MS.
“There may be minor symptoms and disturbances noted in the clinical record that may not be viewed in the context of an MS diagnosis, but may be important information if it can be linked and connected with other data points and general biomedical knowledge.”—DR. SERGIO BARANZINI
“There may be minor symptoms and disturbances noted in the clinical record that may not be viewed in the context of an MS diagnosis but may be important information if it can be linked and connected with other data points and general biomedical knowledge,” said Dr. Baranzini, who helped create a knowledge engine AI/ML model called SPOKE, which brings massive amounts of heterogeneous information (cellular, genetics, pharmacologic, and imaging) into a single graph that can be combined with ML tools to search for connections that underly MS or other diseases.
When applied to large, complex datasets, “AI has the capacity to improve imaging speed and quality, provide new biomarkers, and potentially even to find latent subtypes of MS and clusters of disease, which might lead us to better understand how the disease arises and progresses,” said Michael Dwyer, PhD, associate professor of neurology and biomedical informatics and director of neuroinformatics development at University at Buffalo, who uses ML methods to look for MS biomarkers on MRI.
“With MS, we have a lot of smoking guns, but we still don’t know where this disease is coming from, and AI is capable of potentially looking in more directions at once than we are.”
However, Dr. Dwyer stressed that “a healthy dose of caution” is needed as MS experts figure out ways to translate AI/ML research into validated and useful clinical tools. At this point, “although things are improving, AI is still largely a black box,” he said, “sometimes making the right predictions but unable to explain the reasoning behind its success.”
MRI Scans Might Reveal More
Researchers believe much of the potential for AI/ML technology for MS lies in brain imaging. Dr. Dwyer focuses on thalamic atrophy, which is associated with MS-related neurodegeneration, but the dilemma is that there is no simple, reliable way to measure thalamic volume outside of controlled research settings.
“The things we are looking for are very subtle changes,” said Dr. Dwyer, whose team is using an AI approach called the DeepGRAI tool. “When we are looking at brain atrophy, we are talking about changes of one-tenth of one percent.”
He said ML algorithms also are being tested to identify novel markers of atrophy, which could be used along with known markers and clinical factors to make an MS diagnosis.
Currently, doctors often rely on qualitative judgment in determining if they see or don’t see a new lesion, Dr. Dwyer said. “AI/ML tools will allow for more quantitative-based diagnosis and prognosis, which might be able to be achieved sooner and with a high degree of certainty,” he said. “Right now, even with the results of a high-quality MRI scan, some MS diagnoses are still equivocal, and prognosis is extremely difficult.”
Dr. Dwyer said that ideally there would one day be AI/ML models that could link all relevant MRI findings and other test results—white matter disease burden, thalamic atrophy, cortical atrophy, spinal cord atrophy, fluid biomarkers—into a single diagnostic panel to determine a patient’s MS status.
Dr. Dwyer said medical researchers often face the challenge of translating findings from carefully controlled, high-cost, research setting into everyday clinical care, and that will be true for AI/ML.
“These AI/ML algorithms are very data hungry,” Dr. Dwyer said, “and to make the results relevant and broadly applicable to the MS community, the data being fed must come from a broad spectrum of patients, not just from one MS center or registry. Another challenge is to get longitudinal, up-to-date patient data incorporated into [an] AI/MS model.”
Dr. Dwyer is a member of the AI/ML working group of the North American Imaging in MS Cooperative. He and his colleagues hope that more centers can work together and pool data using techniques like federated learning to overcome these problems.
Some health care technology companies already use lessons learned from ML-related research to develop MR machines that make it possible to do shorter but still revelatory scans, with fewer imaging slices collected, he noted. A key goal would be “the ability to scan someone in five minutes,” he said, because even the more limited number of images could be interpreted using Ai/ML applications, and such short scans would dramatically improve the availability and accessibility of MRI.”
The Eyes May Be Telling
AI/ML technology also could help reveal neurodegenerative changes in the eye indicative of MS, perhaps well before other signs become obvious.
Rachel Kenny, PhD, assistant professor of neurology and population health at NYU Grossman School of Medicine, is part of a research team developing and testing ML algorithms to determine whether data from a simple eye test—optical coherence tomography (OCT)—could be useful for diagnosing MS.
OCT is an imaging method that uses light waves to take cross-sectional pictures of the back of the eye, including the retina and optic nerve. “It is a very simple test,” Dr. Kenney said, and often part of ophthalmologic care to check for signs of glaucoma, diabetic retinopathy, and macular degeneration.
Optic neuritis, a swelling of the optic nerve, for example, is associated with MS, though other conditions besides MS can cause the problem.
Dr. Kenney said that “neurodegeneration that is happening in the brain is also happening in the eye.” She and other researchers use ML methods to study whether images captured by OCT might reveal signs that could predict the onset of MS symptoms.
In a study published in Neurology in September 2022, her team used an ML algorithm to analyze OCT results in connection with MS. The researchers found that OCT results were a reliable prediction tool, though they were most reliable for diagnosing MS when used in combination with data from other sources, such as vision testing. The research team uses ML to answer other MS questions as well.
“We are using machine learning to distinguish different demyelinating disease types, including forms of inflammatory optic neuropathy,” said Dr. Kenney.
A poster featuring some of the findings was presented at the ECTRIMS-ACTRIMS (European Committee for Research and Treatment in Multiple Sclerosis-Americas Committee for Treatment and Research in Multiple Sclerosis) Joint Meeting in Milan in October.
Dr. Kenney said she and her colleagues are also interested in exploring whether OCT could be an especially useful diagnostic tool for MS in middle and lower-income countries where access to more expensive MRI may not be as available as in higher-income countries.
“The benefit is that patients could potentially begin to get treated early, and the MS treatments are so good that they have the potential to prevent disability,” she said.
Looking for Brain-Gut Connections
AI/ML tools also are helping to advance research on the brain-gut connection, an important area of MS research. Dr. Baranzini said researchers traditionally use data sets to pose specific questions they want answered—for example, “Does exposure to X increase the risk of getting Z?” AI/ML technology could potentially integrate and analyze millions of data points in parallel, he said, and make connections that perhaps researchers hadn’t even thought of. With a disease as complex and heterogeneous as MS, AI/MI could be particularly helpful in achieving early and accurate diagnosis, he added.
“One of the challenges has been the availability of large, high-quality datasets with subjects representative of the population to which the model will later be applied.”—DR. RANGA SAMPATH
“AI has the capacity to improve imaging speed and quality, provide new biomarkers, and potentially even to find latent subtypes of MS and clusters of disease, which might lead us to better understand how the disease arises and progresses. With MS, we have a lot of smoking guns, but we still don’t know where this disease is coming from, and AI is capable of potentially looking in more directions at once than we are.”—DR. MICHAEL DWYER
Dr. Baranzini leads the International Multiple Sclerosis Microbiome Study, which compared the gut microbiome of 576 patients with MS with that of 1,152 non-MS genetically unrelated household controls. The analysis found that patients with MS had significantly different gut bacteria than their healthy counterparts. The analysis also identified biomarkers indicative of functional response to MS therapy, which could be useful in treatment management following diagnosis.
“We can use a machine-learning algorithm to extract from the data the most likely neighborhood of information that corresponds to the patient and then use that to try to classify patients with a particular profile as likely to develop MS or not MS,” he said.
“The data generated in this project can be integrated into SPOKE and feed ML models to identify complex and likely predictive patterns in MS patients,” he said.
“MS, I think, is still a mystery in many regards,” Dr. Baranzini said, but with AI/ML tools, “we will be able to integrate much more information on the biological, immunological, genetic, and imaging aspects, and altogether that will help us understand MS at a much deeper level.”
Other Challenges Ahead
Mitchell Wallin, MD, MPH, FAAN, associate professor of neurology at University of Maryland and director of the VA MS Center of Excellence–East, said that AI-ML-informed technology could help address some of the limitations that exist in MS care. New diagnostics based on known or novel biomarkers would be welcome tools since about 30 to 40 percent of people diagnosed with MS present with non-classic symptoms, and even when an MRI is done for suspicion of MS, the results can be equivocal, he said.
“We have multiple therapies for controlling relapsing-remitting MS,” Dr. Wallin said. “Treating and understanding progressive MS is really a challenge.”
Another point in MS care that needs improving is understanding which patients will most likely benefit from which therapy, Dr. Wallin said.
“AI/ML algorithms might help reveal specific MS phenotypes and patient comorbidities that should be prioritized when making therapeutic decisions and “where you might have better outcomes over time,” he added, and AI/ML-findings also might help advance the development of neuroprotection therapies for MS.
Ranga Sampath, PhD, who heads the Center for Innovation in Diagnostics at Siemens Healthineers in Tarrytown, NY, said the growth of AI/ML technology in health care will require building reliable datasets that come from varied sources.
“One of the challenges has been the availability of large, high-quality datasets with subjects representative of the population to which the model will later be applied,” said Dr. Sampath, whose company is developing an AI model that uses routine blood biomarker measurements over a period of time combined with other demographics and clinical information to predict the risk of MS.
The model is based on thousands of data points from electronic health records of 3,000 patients (some of whom had MS) in the MIMIC-IV dataset from Beth Israel Deaconess Medical Center.
Siemens Healthineers researchers recently reported early findings at a meeting of the Association for Diagnostics and Laboratory Medicine that showed that the AI model is highly predictive and could potentially be used to flag patients at risk for MS years before symptoms become apparent.
Dr. Sampath is optimistic about the future of AI for MS and health care in general. “Critical answers to improving patient care could be at our fingertips,” he said.
Disclosures
Dr. Baranzini has received consulting fees from EMD Serono. Drs. Kenney and Wallin had no disclosures. Dr. Dwyer received grant support from Novartis, Bristol Myers Squibb, Mapi Pharma, Merck Serono, Keystone Heart Ltd., Protembis GmbH, and V-Wave Ltd. and consulting fees from Bristol Myers Squibb, Merck Serono, and Keystone Heart Ltd.
