Subscribe to eTOC

Artificial Intelligence Program Provides Rapid, Accurate Diagnosis of Dystonia

Article In Brief

Researchers were able to distinguish and identify isolated focal dystonias with a great deal of accuracy using a deep learning method of artificial intelligence.

Figure

In a fraction of a second, the DystoniaNet platform analyzed raw MRI to diagnose dystonia.

Using a “deep learning” method of artificial intelligence (AI), researchers were able to distinguish and identify isolated focal dystonias with a high degree of accuracy, according to a paper published on September 28 in Proceedings of the National Academy of the Sciences.

The method incorporated an algorithm based on an analysis of microstructural neural networks in brain MRI images, and proved 98.8 percent accurate in distinguishing three types of clinically diagnosed dystonias from healthy controls.

The AI diagnostic platform, called DystoniaNet, enabled the research team to reach the same diagnostic conclusion as two leading clinicians who specialize in dystonia—but in contrast to the typically long process it usually takes to diagnose the disorder, the AI method took only 0.036 seconds to diagnose the disorder.

Typically, diagnosing dystonia is a challenging process, the study authors pointed out. “At present, there is no biomarker or gold standard diagnostic test for dystonia, and there are no current technologies or approaches that address this problem,” they wrote. “We demonstrate that our automatically defined microstructural neural network biomarker, together with its algorithmic platform, DystoniaNet, provides objective, accurate, fast, and cost-efficient diagnosis of isolated focal dystonia.”

The senior author of the paper told Neurology Today that she is in talks with four academic centers to begin validation studies.

“We need to bring this into clinical settings to have physicians evaluate the technology with many more patients and types of dystonia,” said Kristina Simonyan, MD, PhD, Drmed, associate professor of otolaryngology—head and neck surgery at Harvard Medical School and director of laryngology research at Massachusetts Eye and Ear.

She emphasized that, like any other test, DystoniaNet is not meant to replace clinicians; rather it aims to enhance their diagnostic decision-making.

“The idea is to give neurologists a clinically valuable tool they can use for additional reference,” Dr. Simonyan said. “It would make no sense to dispense with a clinical assessment or case history.”

Three leading neurologists who specialize in movement disorders, including dystonia, called the paper exciting and even revolutionary, but said its results need to be replicated on larger and more diverse samples of patients.

“This study is exciting—it has a lot of potential,” said Richard L. Barbano, MD, PhD, FAAN, professor of neurology and chief of the movement disorders division at the University of Rochester, who was not involved with the research. “But they are comparing people who have been diagnosed by expert clinicians as having dystonia against perfectly healthy controls. How good is their program at diagnosing milder cases, or look-alikes due to some other problem? That's what we really need. My gut tells me they're onto something, but they need to do more work.”

Study Details

To develop the AI platform, Dr. Simonyan gathered raw structural brain images from 612 subjects, including 392 patients with three forms of dystonia—279 with laryngeal, 59, cervical, 54, blepharospasm—and 220 healthy controls. All patients' diagnoses were confirmed by at least two dystonia specialists.

The images were then analyzed by DystoniaNet. Unlike so-called shallow machine learning programs, which take a single run through a data set, DystoniaNet uses a multi-step mathematical function called convolution. Previously used to develop self-driving cars and robotic vision, the method essentially puts each pixel or voxel through a series of filters. After a first pass to find signals at the highest level of magnification, the system takes a second and third pass, each time looking at slightly lower levels of magnification—to be certain that initial signals still hold up.

Developed by Davide Valeriani, PhD, a computational neuroscientist and postdoctoral research fellow at Massachusetts Eye and Ear, the DystoniaNet program is given no indication what to look for. It simply searches for any patterns or signals that separates patients from controls, said Dr. Valerian, who was the first author of the study.

Beginning with a training set of 160 patients with laryngeal dystonia and 160 healthy controls —and a validation set of 60 patients and 60 controls, DystoniaNet identified significant areas in the corpus callosum, thalamic radiation, inferior fronto-occipital fasciculus and temporal gyrus, “all of which have been previously reported to be abnormal in patients with various forms of dystonia,” the paper noted.

“Alterations in these regions are thought to contribute to abnormal heteromodal sensorimotor processing and executive control of goal-oriented motor behaviors in patients with isolated dystonia,” the authors stated.

To check the diagnostic generalizability of the findings, Drs. Simonyan and Valeriani then ran the program on a second group of 172 patients with three different forms of isolated focal dystonia. The overall accuracy in those three groups, 98.8 percent, was such that DystoniaNet concluded that only six of the patients had to be referred for further examination.

The performance of the method did not depend on any particular type or setting of MRI. The findings “remained stable independent of the magnetic field strength (accuracy range: 98.0 to 100%), MRI scanner vendor (accuracy range: 96.9 to 100%), head coil (accuracy range: 95.2 to 100%), T1-weighted image acquisition sequence (accuracy range: 98.3 to 100%), or a data collection site (accuracy range: 97.6 to 100%).”

In contrast to prior studies that have focused on gray-matter changes associated with dystonia, the DystoniaNet method found that “white matter alterations across different forms of dystonia emerge as a more common feature of this disorder.”

The abnormalities detected by DystoniaNet, Dr. Valeriani said, are “structural differences in how these regions are shaped. We are not able to say at this stage whether this area is, for instance, two inches rather than one and a half inches. But we are able to say that there is some sort of change in the structure of this area that is useful enough to make a diagnosis. Most importantly, these changes in the microstructure of the brain are too small or subtle to be seen by the naked eye.”

Dr. Valeriani previously presented the study at MDS Virtual Congress 2020, the virtual meeting of the International Parkinson and Movement Disorder Society. DystoniaNet is a patent-pending proprietary platform developed by Dr. Simonyan and Dr. Valeriani in conjunction with Mass General Brigham Innovation.

Expert Commentary

Dr. Barbano welcomed the study's finding with this caveat: It is not clear what DystoniaNet is picking up in the brain areas it identifies as biomarkers.

“I don't know whether what they're picking up is causing the dystonia or is a result of the person having abnormal movements for many years,” Dr. Barbano said.

David Eidelberg, MD, FAAN, professor of neurology and director of the Center for Neurosciences at The Feinstein Institute for Medical Research in Manhasset, NY, has published multiple papers describing automated algorithms for the differential diagnosis of parkinsonism based on analysis of MRIs. But while his algorithm was trained specifically to search for variations in metabolic networks, he noted, DystoniaNet was designed to look for any patterns it could link to the diagnosis of dystonia, free of constraints.

“In a way, it's revolutionary,” Dr. Eidelberg said. “It searches without a premise to do pattern recognition. It's totally naive, yet it appears that they were successful.”

He emphasized that much more validation will be necessary before DystoniaNet is embraced by clinicians. But, he added, “The goals of this study were well thought out. It was done by good people with very concise endpoints. It has all the ingredients for being a useful application in the future.”

For now, the study was able to demonstrate only that DystoniaNet can distinguish three of the hundred or so types of dystonia from healthy controls, said Mark Hallett, MD, NIH Distinguished Investigator in the Human Motor Control Section of the National Institute of Neurological Disorders and Stroke.

“The data look good for what has been studied,” said Dr. Hallett, who has specialized in the study of dystonia. “That's why it needs to be investigated further to see how generalized it can be.”

When he began his research, he said, “Dystonia was really not recognized at all. I remember talking with other neurologists about things like writer's cramp, and nobody believing me that it was dystonia. Things have changed in that regard. However, it's still rare enough that once you get outside of neurology, it's difficult for many primary care providers to recognize.”

He hopes to collaborate with Dr. Simonyan and other investigators in the near future on a larger study, Dr. Hallett said.

Disclosures

Dr. Simonyan has received consulting fees from Jazz Pharmaceuticals, Inc. Dr. Valeriani has received payment for travel-related expenses from Amazon Web Services for attending re:invent 2018. Dr. Barbano has received consulting fees from Allergan and Merz and research funding from Revance. Dr. Eidelberg serves on the scientific advisory board and receives honoraria from the Michael J. Fox Foundation for Parkinson's Research; serves on the scientific advisory board and receives personal fees from Ovid Therapeutics; and receives consulting fees from MeiraGTx. Dr. Hallett had no relevant disclosures.

Link Up for More Information

• Valeriani D, Simonyan K. A novel microstructural neural network biomarker for dystonia diagnosis identified by a DystoniaNet deep learning platform https://www.pnas.org/content/early/2020/09/30/2009165117. Proc Natl Acad Sci 2020; Epub 2020 Sept 28.
    • Schirinzi T, Sciamanna G, Mercuri NB, et al. Dystonia as a network disorder: A concept in evolution https://journals.lww.com/co-neurology/Abstract/2018/08000/Dystonia_as_a_network_disorder__a_concept_in.20.aspx. Curr Opin Neurol 2018; 31: 498–503.
    • Topol EJ. High-performance medicine: The convergence of human and artificial intelligence https://www.nature.com/articles/s41591-018-0300-7. Nat Med 2019; 25:44–56.
    • Battistella G, Simonyan K. Top-down alteration of functional connectivity within the sensorimotor network in focal dystonia https://n.neurology.org/content/92/16/e1843. Neurology 2019;92, e1843–e1851.