Graduate students at Stevens Institute of Technology in New Jersey believe they have found an innovative way to detect Parkinson’s disease in its early stages – an otherwise difficult and expensive undertaking.
“We are confident someone could use our methods to implement low-cost movement disorder detection," says graduate cybersecurity student Divyendra Patil, who teamed with fellow cybersecurity student Rahul Yadav to develop their Park-Detect concept that won a top prize at a recent hackathon competition at Stony Brook University in New York.
The Center of Excellence in Wireless and Information Technology (CEWIT) hosted the inaugural Hack@CEWIT, an interdisciplinary student hackathon focusing on industry-relevant internet of things (IoT) and microservices challenges to advance enterprising, student-powered software solutions in the Center’s next-generation research and education facility on the Long Island, New York campus Feb. 17-19.
To build their system, the duo — assisted by graduate students Sagar Jain and Poornima Pundir — relied upon a public database of keystroke data created from more than 200 international subjects, including some confirmed with Parkinson's disease and some healthy. Participants installed temporary “keylogger” applications that automatically tracked and recorded their keystrokes as they worked on computers.
With that pool of keystroke data in hand, Patil and Yadav then added an inexpensive camera to the mix. By breaking typing videos into frames and programming a convolutional data network — a series of repeated mathematical process that filter and classify data – their system learned to figure out which hand a subject was using to press each key. They also ran Python language codes on the data to filter further by hand, key, stroke duration, time lapse and relative direction to the next keystroke.
Finally, armed with their analyses, the pair ran repeated simulations, trying to predict Parkinson's cases with no other prior knowledge of a person except his or her age and typing results.
After training the network for just two days, Patil and Yadav found they could correctly detect a confirmed Parkinson's patient in a case pulled blindly from the Australian dataset an impressive 72 percent of the time.
"We felt this idea might work because typing response time is known to be slower and less accurate in Parkinson's sufferers," explained Patil. "The average person might type, let's say, 70 or 80 words per minute; the average Parkinson's patient can't type more than 40 words per minute,” he said.
According to a Stevens Institute release March 26, the researchers said the next step would be to bring in facial reactions to see what the user is more concentrated on as well as gait recognition.
Without financial backing to develop the idea further, Patil and Yadav both hope to continue to pursue the project during summer internships or after graduation.
"The Park-Detect methodology could also be used to detect other movement disorders, with modifications and by understanding the various symptoms of those other diseases," said Patil. "We are not doing this to make money. We want to help people."