Applications of artificial intelligence (AI) are becoming more and more common in daily life, and medicine. It’s public knowledge that AI algorithms are much better at detecting a pneumonia on a chest X-ray than radiologists of flesh, blood and … insomnia. A similar problem arises when judging whether MS lesions on a brain scan have increased in volume over time. (Un)fortunately, lesions never double in volume. Radiologists need to pick up volumetric increases of 1 or 2 mL and it’s unrealistic to expect that this will be done flawless. Hence, a clear window for AI algorithms to improve healthcare and patient outcomes. The AI algorithms recognise patterns in data, and create their own logic. To gain useful AI insights and predictions, it is key that the models are trained using extensive amounts of input data, far beyond what any brain could handle.
A new study applied AI on brain MRI data to categorise pwMS in different subtypes. They used input data from 6322 pwMS to define MRI-based subtypes on an independent cohort of 3068 pwMS for validation. Scans were processed to quantify volume loss and tissue damage in different brain regions and MS lesion volume. These data were used to train and validate the AI algorithm, and three MRI patterns became apparent: a cortex-led, a white matter-led, and a lesion-led subtype. The cortex-led subtype with conspicuous damage in the outer layer of the brain was most frequent. People with the lesion-led subtype had the most aggressive disease course: more contrast-enhancing lesions, more relapses and a higher risk of accumulating disability over time. Importantly, people with the lesion-led MS subtype showed a positive treatment response whereas people with the cortex-led subtype did not respond well to MS treatments.
PwMS are wondering now: which MRI subtype am I?
Very good question. However, I’m not able to give you an answer. These AI algorithms are not battle-zone ready. This means that we cannot apply them in clinical practice (yet). The current algorithm was trained with MRI data from the ‘perfect’ world: pharma-led clinical trials. In reality, there is much more technical noise and more variable intervals between first symptoms-first scan, and first scan-second scan, etc. However, this is probably the least unsurmountable problem as logistics are – as always- only a temporary distraction.
Therefore, the better question would be: does it matter which subtype am I?
Maybe. The AI algorithm clearly showed that response to immunotherapy is largely lesion-dependent. This is very interesting but also largely in line with our current – admittedly simplistic – non-AI-based clinical practice when it comes to treating pwMS. We select pwMS for treatments based on the presence or absence of new/enhancing lesions on their latest brain scan. Is it excluded that cortex-led pwMS develop new lesions? No, not at all. Would anyone refuse treatment in a patient with a cortex-led subtype who now has a new lesion? No. Because essentially every pwMS is a bit over everything. Everyone is a bit of lesion-led, a bit of cortex-led and a bit of white matter-led. Although you match better with one of these subtypes, you are unlikely to be a perfect match. The results in the paper clearly show that the lesion-led and cortex-led subtype are two extremes of a spectrum, with the white-matter led subtype sitting right in the middle. This means that annual relapse rate is higher in the lesion-led > white-matter-led > cortex-led subtype. Therapeutic response followed the same pattern of gradual decay. This clearly implies that the AI algorithm categorised an MRI spectrum in three boxes; like splitting a circle into three.
Nonetheless, I am pretty convinced that these AI boxes are more accurate than the clinical boxes neurologists like to use (relapsing-remitting vs.progressive). Clinical phenotyping is outdated, and the only thing that probably matters when starting MS treatments is whether you have ongoing/active inflammation or not. Luckily, pharmacological walls are progressively breaking down with the advent of Ocrevus for primary progressive MS and siponimod for secondary progressive MS. And if AI algorithms are gonna knock on this wall, we are happy to join. Knock Knock!
Disclaimer: Please note that the opinions expressed here are those of Ide Smets and do not necessarily reflect the position of the Barts and The London School of Medicine and Dentistry nor Barts Health NHS Trust.
Nat Commun 2021 Apr 6;12(1):2078. doi: 10.1038/s41467-021-22265-2.
Arman Eshaghi 1 2, Alexandra L Young 3 4, Peter A Wijeratne 3, Ferran Prados 5 3 6, Douglas L Arnold 7, Sridar Narayanan 7, Charles R G Guttmann 8, Frederik Barkhof 5 3 9 10, Daniel C Alexander 3, Alan J Thompson 5, Declan Chard # 5 11, Olga Ciccarelli # 5 11
- PMID: 33824310
- PMCID: PMC8024377
- DOI: 10.1038/s41467-021-22265-2
Multiple sclerosis (MS) can be divided into four phenotypes based on clinical evolution. The pathophysiological boundaries of these phenotypes are unclear, limiting treatment stratification. Machine learning can identify groups with similar features using multidimensional data. Here, to classify MS subtypes based on pathological features, we apply unsupervised machine learning to brain MRI scans acquired in previously published studies. We use a training dataset from 6322 MS patients to define MRI-based subtypes and an independent cohort of 3068 patients for validation. Based on the earliest abnormalities, we define MS subtypes as cortex-led, normal-appearing white matter-led, and lesion-led. People with the lesion-led subtype have the highest risk of confirmed disability progression (CDP) and the highest relapse rate. People with the lesion-led MS subtype show positive treatment response in selected clinical trials. Our findings suggest that MRI-based subtypes predict MS disability progression and response to treatment and may be used to define groups of patients in interventional trials.