Novel AI-based platform diagnoses dystonia from MRI scans

Reuters Health - 28/09/2020 - An artificial intelligence (AI)-based deep learning platform, DystoniaNet, diagnosed the neurological disorder from MRI scans with close to 99% accuracy, researchers say.

DystoniaNet is a patent-pending proprietary platform developed by Drs. Kristina Simonyan and Davide Valeriani of Massachusetts Eye and Ear Infirmary in Boston in conjunction with Mass General Brigham Innovation.

"DystoniaNet was not developed to replace clinicians, but rather as an objective test to reduce time-to-correct diagnosis," Dr. Simonyan told Reuters Health by email. "There are currently no biomarkers of dystonia and consequently, no gold standard tests. Our study identified a microstructural neural network biomarker that is reflective of dystonia pathophysiology."

"The availability of DystoniaNet as a computerized diagnostic algorithm will allow providers to conduct diagnostic screenings not only in the clinic but also via growing telemedicine channels," she said. "The accessibility via the cloud-based platform will enable its use from any location worldwide."

"Currently, the DystoniaNet prototype is available as an experimental diagnostic platform," she noted. "We are discussing its implementation for dystonia diagnostics in the neurology and laryngology clinics within Mass General Brigham, as well as other centers nationwide. In parallel, we are planning to conduct a larger, multi-center clinical validation study, as part of which we will pursue biomarker qualification with the FDA. We are committed to accelerating its full clinical implementation within the next few years."

As reported in Proceedings of the National Academy of Sciences of the United States of America (PNAS), the researchers compared brain MRIs of 392 patients with three forms of isolated focal dystonia - laryngeal (spasmodic), cervical, and blepharospasm - and 220 controls without dystonia. Participants' mean age was about 49 and 64% were women.

DystoniaNet identified clusters in the corpus callosum, anterior and posterior thalamic radiations, inferior frontooccipital fasciculus, and inferior temporal and superior orbital gyri as biomarker components. These brain regions are known to contribute to abnormal interhemispheric information transfer, heteromodal sensorimotor processing, and executive control of motor commands in dystonia.

The biomarker demonstrated an overall accuracy of 98.8% in 0.36 seconds per person in diagnosing dystonia, with referral of 3.5% of cases due to diagnostic uncertainty.

"DystoniaNet significantly outperformed shallow machine-learning algorithms in benchmark comparisons, showing nearly a 20% increase in its diagnostic performance," the authors write. Further, the biomarker demonstrated a "substantial" improvement over "the recently reported 34% diagnostic agreement rate between clinicians with different expertise based on extensive diagnostic workup and syndromic approach," they note in the report.

Dr. Simonyan said, "Next steps should involve an implementation of the identified biomarker and its DystoniaNet diagnostic platform in a clinical setting, where physicians have an opportunity to test it in prospective patients."

"This is an extremely interesting and important approach. The diagnosis of dystonia remains clinical, and this poses numerous problems. Often clinicians do not agree on the criteria, and there are disorders that mimic dystonia but are not dystonia," Dr. Steven Frucht, co-director of the Fresco Institute for Parkinson's and Movement Disorders at NYU Langone Health in New York City, commented by email to Reuters Health.

"Inclusion of patients in therapeutic and disease-modifying trials is also an issue that is hampered by lack of objective criteria for diagnosis," he noted.

"Validation of the DystoniaNet pattern in a larger cohort, in different countries, with different forms of dystonia" is needed, Dr. Frucht added. "Potential obstacles are not high in terms of obtaining routine MRI imaging. Access to the program, and implementation in resource-challenged areas, are obstacles that need to be addressed."

By Marilynn Larkin

SOURCE: https://bit.ly/3kWcNkO PNAS, online September 28, 2020.



Posted on 28 September 202022 October 2020