AI identifies new multiple sclerosis subtypes

AI identifies new MS subtypes

Date: 9th April 2021

Over 2.8 million people globally are living with multiple sclerosis (MS), a lifelong condition where the body’s own immune system attacks the central nervous system (CNS), damaging the protective myelin sheath that covers nerve fibres.  MS can be categorised into four phenotypes based on clinical evolution, defined as either relapsing or progressive.  Now, researchers have used artificial intelligence (AI) on MRI brain scans to classify MS subtypes based on pathological features, identifying three new subtypes which will better guide effective treatment choice and progression risks.

There are currently two descriptors of the four existing MS phenotypes, the first -disease activity which is determined by relapses or new activity evident by MRI, and the second – progression of disability.  Together these are routinely used to select patients and to guide treatments in clinical trials.  However, there is an argument that redefining disease subtypes based on biology rather than clinical criteria alone would be more informative, and would better place patients most likely to benefit from investigational medication.

Now, researchers from University College London, UK, have used training MRI datasets from MS patients to define MRI-based subtypes which were then validated on an independent cohort.  They identified three new MS subtypes based on the earliest abnormalities and found that they predicted MS disability progression and response to treatment.

There are many similarities between the existing MS phenotypes and patients can transition from one to another, often without knowing.  The clinical data is largely subjective, relying on a patient’s recall of symptoms, and changes in MRI features and disease pathology, which are hard to distinguish.  The team therefore, wanted to a data-driven solution to uncover disease subtypes.  So they used their recently developed unsupervised machine learning algorithm, called Subtype and Staging Inference, SuStaIn, to reveal distinct disease subtypes with distinct temporal progression patterns.

SuStain was used on a training dataset of MRI scans from 6322 MS patients, and uncovered three previously unknown MS subtypes with distinct patterns of MRI abnormalities.  These were named cortex-led, normal-appearing white matter-led, and lesion-led.  In a separate independent cohort consisting of 3068 patients these subtypes were then validated.

The team then wanted to determine whether there were differences in disease activity, progression and treatment response between the newly identified subtypes.  By examining the clinical characteristics of each subtype they found the patients with lesion-led subtype had the highest Expanded Disability Status Scale (EDSS) meaning they had the highest risk (>30%) of disability progression, the largest baseline lesion load, and the highest rate of relapse. However, they were also the group that showed positive treatment responses in selected clinical trials.  The most frequent subtype was cortex-led, and age and sex had no bearing of MS subtype. Together, these results suggested that the new MRI-based subtypes could predict MS disability progression and response to treatment.

Conclusion and future applications

The team here have presented a data-driven, AI-discovery of three new MS subtypes, characterised by distinct temporal patterns of MRI changes.  The results suggest that MRI is strong candidate for AI reclassification, as it better reflects MS pathogenesis than the current indicators, which are based purely on clinical descriptions.   SuStain subtypes were predictive of an individual’s risk of disease progression and were also associated with response to treatment.

This groundbreaking discovery has the potential to be a real game-changer, and would be a valuable clinical tool for informing specialists of both disease evolution and selection of patients for clinical trials.

Although clinical trials of SuStain will first be required, it is an important milestone in predicting individual responses to treatment.  As more novel therapies hit clinical trials and new datasets are acquired, SuStain can evolve alongside, perhaps identifying yet more subtypes and defining groups of patients assigned to the most efficacious interventional trials.  The team are now extending SuStain’s approach by including other clinical information, with the aim of precision medicine – where an individual patient will have the right treatment, at the right time predicted by AI.

The AI field is currently moving at an astonishing rate and this works adds to ongoing efforts of using AI to drive diagnostics and predict disease outcomes.  We have seen AI being leveraged to predict mortality from echocardiograms, to diagnose autism from maternal biomarkers and early-stage breast cancer, as well as predicting  dynamics of brain networks in response to microstimulation. AI is starting to have a high impact in the medical and healthcare field, and the reclassification of many disease types, especially those associated with high-resolution imaging, will accelerate precision medicine and improve patient care and outcomes.


For more information please see the press release from UCL


Eshaghi, A., A. L. Young, P. A. Wijeratne, F. Prados, D. L. Arnold, S. Narayanan, C. R. G. Guttmann, F. Barkhof, D. C. Alexander, A. J. Thompson, D. Chard and O. Ciccarelli (2021). “Identifying multiple sclerosis subtypes using unsupervised machine learning and MRI data.” Nature Communications 12(1): 2078.