Amazon Research Award recipient develops new tool to diagnose dystonia

Imagine waiting five years to get a diagnosis. That is the reality facing those who may be battling dystonia, a potentially debilitating neurological condition that requires a painstaking process to identify. Dystonia is characterized by involuntary muscle movements that can manifest throughout the body, but usually show up in one area. Commonly, eyelid muscles, neck muscles, hands, or vocal cords are affected—all of which can seriously interfere with people’s lives. For example, laryngeal dystonia, which impacts the voice, can make speech difficult or impossible.

Symptoms often first appear in midlife, when patients are hitting career highs and family pressures mount. “For patients it’s just really scary,” says Dr. Kristina Simonyan, a neuroscientist and the head of the Dystonia and Speech Motor Control Laboratory at Massachusetts Eye and Ear and Harvard Medical School. “It’s one thing if you know what your issue is—but not to know is very difficult in any health diagnosis.”

The long and difficult diagnosis time, which takes an average of five-and-a-half years for a typical patient, is exactly why her laboratory spent the last decade-plus developing DystoniaNet, a new AI-based deep learning platform that can perform the task in a fraction of a second.

Dystonia is considered a rare disorder; it affects about 300,000 people in the United States. It’s also often undiagnosed or misdiagnosed, so the actual numbers may be higher. Why is it so tough to know if someone has it? There’s no biomarker or gold-standard diagnostic test for dystonia, so doctors have to go through a process of elimination to determine whether the symptoms suggest dystonia or another neurological issue, like Parkinson’s. Even that process isn’t straightforward: “The clinical criteria are vague, and they’re not standardized. They depend on the clinician’s experience and expertise,” Simonyan says.

Symptoms also vary between patients—and even within the same patient—and can fluctuate over time. “If [a patient] sees the clinician in the morning and then sees another clinician that afternoon, there could be a discrepancy in opinions, because symptoms also change at different times of the day and different days within the week,” says Simonyan.

Those challenges are exactly why a dystonia diagnostic test is so important to develop.

Simonyan’s work on dystonia began in 2004, when she was a postdoctoral researcher with the National Institutes of Health. Her interest in laryngology, which began in medical school when she became fascinated by hearing and voice production, led her to focus on neural control of voice production and then laryngeal dystonia—but she was starting from scratch.

“There was really nothing known,” Simonyan says. “At that time, there weren’t any neuroimaging studies done to see where the abnormalities are, what is normal or abnormal, or how voice and speech are processed and output by the brain in these patients.”

She published paper after paper on research aimed at figuring out the functional and structural abnormalities in the brains of patients with laryngeal dystonia, as well as the connectivity between brain regions and speech production. “For many, many years, we were trying to understand the pathophysiology of this disorder in order to identify better diagnostic and better treatment options,” says Simonyan.

Four years ago, all that work started to pay off.

For patients with dystonia, an MRI usually doesn’t show an abnormality specific to the disease, says Simonyan. But over time, radiological images acquired for research studies consistently picked up hallmarks of dystonia within the brain—microstructural changes that a clinical MRI wouldn’t show. “We capitalized on that,” she says. 

In 2016, Simonyan and her research team published their first paper using machine learning linear discriminant analysis (LDA). LDA is a method used to classify a set of data into distinct groups—in this case the structural and functional MRI neuroimaging of different phenotypes and genotypes of dystonia. Once those markers were identified and categorized from a large number of patients, they were fed into a machine learning algorithm.

The goal, Simonyan says, was to see whether images from patients already diagnosed with dystonia could be used to classify undiagnosed patients who might have the disease. It worked, with a success rate of 81% correct diagnoses. That’s better than the current rate among physicians (about 34%), but still not good enough to move to clinical setting—and it meant doctors treating patients needed to learn how to do time-consuming image analysis and other tasks to run the program.

Time in the clinic is limited and the busy doctors there have other priorities. Simonyan wanted an even better test. “That was my motivation to turn this research to deep learning,” she says. Her team set out to build a tool that automated everything—not only the machine learning analysis of data, but the data input and processing as well.

Together with her postdoctoral fellow, she did just that over the past two years, increasing the level of automation in several iterations. “With the support of Amazon Web Services, we were fortunate to have access to superb computational resources and combine them with our large data set of patients. The final product was DystoniaNet,” she says.

And while the data were centered on laryngeal dystonia, the test also works to diagnose other forms of dystonia that affect the neck muscles (cervical dystonia) or eyelid muscles (blepharospasm). It has a correct diagnosis rate of 98.8%, and best of all, it doesn’t take years or months: Simonyan’s test takes just .36 seconds.

Simonyan hopes that DystoniaNet will move to clinical practice (expanded testing is underway) and says clinicians and researchers have been very excited at her presentations.

In the meantime, her team continues to refine DystoniaNet. A new capability would incorporate the ability to rule out dystonia and pinpoint other neurological disorders, such as Parkinson’s and essential tremor, making it useful to many more physicians—and patients.

DystoniaNet doesn’t exclude the physician from the process of diagnosis, Simonyan says: “It just helps them and provides an objective diagnostic tool that they have been lacking all this time.”

She says Amazon’s ARA funding for her work on DystoniaNet was critical. Her team was able to hire additional help and it gave them access to the cloud, where they could speed up the process of training and testing the model.

“That really made a difference for us to move forward, especially with the very large number of subjects,” she says.

Computing power made a difference too: The computational framework Simonyan’s team used was implemented on an AWS Deep Learning Amazon Machine Image (AMI), and run on the Amazon Web Services EC2 P2, which Simonyan says wasn’t matched by even the very powerful workstations in her lab. That expanded the computational ability to test, train, and refine different iterations of the model. Without that capability, she says, “our time in this process would have been much, much longer—we probably would have been still testing it,” she says.

Her team’s ability to harness the potential of AWS also factored into the decision to fund her grant.

“Our goal is to support researchers, such as Dr. Simonyan and her team, with infrastructure and tools to accelerate their work through the Amazon Research Awards program,” says Taha A. Kass-Hout, MD, MS, director of machine learning, at Amazon Web Services. “We are thrilled to hear that AWS machine learning tools were able to speed up their groundbreaking research and development of DystoniaNet.”

Even with Simonyan’s work, hurdles remain for dystonia patients. The cause is unknown, there is no cure, and available treatments are minimal. All patients can do is work with their doctor to manage their symptoms. But knowing exactly what disease they have, and that it’s not fatal even though it affects their quality of life, can reduce some of the fear and uncertainty for those who suffer from it.

“They go through this period of time where they don’t know if they’re dying or not, or what’s wrong with them,” says Simonyan. “So all of them are relieved when they get the diagnosis.”