Kristina Simonyan, MD, PhD, DrMed, and Davide Valeriani, PhD, offered insight into the use of DystoniaNet, which identified the condition with 98.8% accuracy in a matter of 0.36 seconds.
In September 2020, data were published from an assessment of an artificial intelligence (AI) based deep learning platform which suggest that it can quickly and accurately identify individuals with dystonia, a notoriously difficult to diagnose disorder using magnetic resonance imaging (MRI) data.
The investigators, Kristina Simonyan, MD, PhD, DrMed, director, Laryngology Research, Mass Eye and Ear; associate neuroscientist, Massachusetts General Hospital; and associate professor, Otolaryngology-Head and Neck Surgery, Harvard Medical School; and Davide Valeriani, PhD, postdoctoral research fellow, Dystonia and Speech Motor Control Laboratory, Mass Eye and Ear and Harvard Medical School, noted that the tool may help clinicians severely cut down the time to diagnosis. The platform identified the condition with 98.8% accuracy in a matter of 0.36 seconds. Specifically, it showed 98.2% accuracy in diagnosing laryngeal dystonia, 100% in diagnosing cervical dystonia, and 98.1% in diagnosing blepharospasm, while referring 3.5% of patients (n = 6) for further examination.
To find out more about DystoniaNet and the assessment, NeurologyLive spoke with Valeriani and Simonyan about what prompted the study and what the potential of the platform might be.
Kirstina Simonyan, MD, PhD: Dystonia is the third most common movement disorder after essential tremor and Parkinson disease, and it affects over 300,000 people in the United States alone and even more worldwide. There is a significant diagnostic challenge that is primarily associated with the absence of a biomarker for dystonia—an objective measure as an indicator of a common process pathophysiological process that happens in this disorder. As a result, there is also no gold standard test for the diagnosis of this disorder, and current diagnostic indications largely remained based on clinical symptom evaluation and syndromic approach.
In addition to the absence of a biomarker and test, the diagnosis is typically impacted by the large variability of symptomatology of this disorder, the circumstances of the evaluation, the state of the patient, the expertise and experience of the clinician, and other non-neurological conditions that can mimic dystonia symptoms.
I have been seeing these challenges along my medical and scientific career, and being trained as a clinician, my focus has been on improving clinical management of this disorder and understanding its pathophysiology and how we can improve diagnostic and therapeutic interventions for these patients. That is a kind of larger picture of how we started thinking about this line of research. There have also been several methodological advances that allowed us to dive deeper into the pathophysiology of the disorder based both on the development of neuroimaging methodologies as well as advanced machine learning algorithms. Bringing this together, we were well-positioned to take on this direction of research.
Davide Valeriani, PhD: What we did was use structural MRI to diagnose dystonia together with a deep learning algorithm. We basically leveraged the advances of deep learning and designed an architecture that was able to look at raw structural MRI and find a biomarker for dystonia that could help with the diagnosis of this disorder.
We specifically took an approach that would be easy to clinically translate, and that's why we used raw structural MRI. This kind of measure is what comes out directly from the scanner and is available to many clinicians in the clinics. What the DystoniaNet really does is take the structural MRI and learn on our training set—we trained this algorithm on a large data set of neuroimaging data that we had available, one of the largest data sets currently available to a single lab. That was what enabled us to use DystoniaNet and to train it, and now what this it does is really learn a biomarker from the data. It's a completely data-driven approach.
We discovered this biomarker that could be used together with the DystoniaNet platform to diagnose dystonia in a fraction of a second because it is very fast to output the diagnosis. It is also able to understand when it's not confident about this diagnosis and referring that patient for further examination. We really developed this over the past few years—the past 2 years—to really be able to easily translate to the clinics and take into account all the feedback we got from the clinicians here.
Kirstina Simonyan, MD, PhD: Some of the findings were not surprising. We would expect that advanced machine learning would be efficient in diagnosing this disorder, based on our prior research. Back in 2016, we published a paper that was among the first papers on machine learning use in dystonia, where we explored analysis-driven, not data-driven like in this case, markers and how well they can perform for diagnosing dystonia. Those results were quite encouraging, where we would have an accuracy between 71% and 81% in correctly diagnosing dystonia. Research from other labs also has shown that similar biomarkers and analysis-driven biomarkers can be implemented for the diagnosis of dystonia.
However, that was not good enough for us. That type of paradigm would have involved substantial data processing, which would have been clinically challenging to do because clinicians just don't have the luxury of several hours per patient to acquire images, process the data, and then receive the output of the machine learning paradigm. In a way, our study was designed to lessen the burden—the time burden—on the clinician and to speed up the diagnosis. In this case, it is a fraction of a second.
A lot of methodological advances and work went into this development. In a way, it was not greatly surprising that we were able to use neuroimaging for diagnosis and for use as a diagnostic biomarker. But it was, in a way, surprising to that we could actually achieve very high accuracy on testing hundreds of subjects with dystonia and without dystonia, and then also retesting it on almost 1500 healthy subjects to validate this paradigm. Again, we're achieving over 96% accuracy in correctly identifying healthy subjects and patients, and that was done in a fraction of a second. We hoped that this would be the case, and we achieved that.
Kirstina Simonyan, MD, PhD: The other aspect of this was that dystonia is still considered the textbook example of basal ganglia disorder. The major focus in understanding the pathophysiology of this disorder has been focused for many years on basal ganglia so it has been a little bit surprising that we did not find any specific basal ganglia regions that would contribute to this high diagnostic accuracy of the algorithm. But looking in from a different perspective, the current state of the knowledge is that dystonias are a network disorder, and while basal ganglia are involved, there are also other multiple cortical and subcortical regions that have equally significant roles in its pathophysiology. So the identification of white matter regions as components of part of a biomarker in a way was both surprising and not. Again, because it is a basal ganglia disorder, you would think there might be a region in basal ganglia. But on another hand, white matter regions pass fibers between different distant areas in the brain, and you would think that abnormalities in these regions would have a wider impact on the organization of brain structure and function in this disorder. In addition to this, all these regions have been reported across different forms of dystonia, to be abnormal, to carry microstructural alterations. In a way, as you start thinking about what this might mean, in terms of biomarker and diagnostic potential, the surprise goes away and more excitement comes up. You can see that the biomarker truly reflects the pathophysiology of dystonia, rather than picks up and learn something that is not really relevant. It is relevant to part of the physiology, and obviously, the timing is excellent to reduce time to diagnosis and aid clinicians in decision making.
Transcript edited for clarity.