New AI data set aims to advance remote Parkinson’s diagnosis
Web-based study tracks motor and memory tasks from home devices
Written by |
- Parkinson's disease is a neurological disorder causing motor and nonmotor symptoms.
- Traditional in-person clinical evaluations present challenges, prompting the need for remote assessment solutions.
- Scientists have compiled a large data set to train and test AI for objective, remote Parkinson's diagnosis.
A team of U.S. scientists has compiled a large data set of remotely collected data from people with and without Parkinson’s disease, which the scientists hope will serve as an important tool for developing strategies to diagnose the disorder remotely using artificial intelligence (AI).
“By pairing task-derived behavioral features with device type and handedness metadata, this dataset provides a benchmark resource for evaluating AI robustness, subgroup performance, and generalizability in remote [Parkinson’s disease] assessment,” the scientists wrote in “RobustPDx: A Structured Web-Based Parkinson’s Assessment Dataset and Benchmark for AI Robustness to Device Type and Handedness,” which was published in Scientific Data.
In-person assessment poses logistical hurdles
Parkinson’s is a neurological disorder that causes motor symptoms such as slowness and tremor, as well as nonmotor symptoms that can range from memory problems to digestive complaints.
Assessments conducted by a clinical expert are currently the gold standard for diagnosing Parkinson’s and tracking the disease’s severity. But conducting an in-person assessment poses notable logistical hurdles, especially for people who don’t live near specialty centers, and the reliance on a human evaluator makes the assessments inherently subjective.
As it’s become more common for people to have computers and smart devices at home, scientists are increasingly exploring whether these devices could be used to monitor Parkinson’s remotely, offering greater convenience and objectivity.
AI is a broad field of computational analysis that uses large amounts of data and advanced algorithms to identify patterns. For AI tools to be effective, they need to be trained and tested on large, well-defined data sets.
Online platform guides users through a series of tasks
In this study, researchers used an online platform to generate a new data set to train AI for Parkinson’s assessments. Essentially, the online platform guided users through a series of tasks — several movement-based assessments performed by moving the mouse or typing on the keyboard, followed by a memory test. From these tasks, the researchers defined dozens of specific variables, such as the precise timing of key clicks and the exact variability of mouse movements.
The researchers noted that the online tasks could be completed by anyone with a computer and internet access.
“The naturalistic, remote design enabled the collection of motor and cognitive performance data under conditions relevant to real world digital-health applications, including variation in device,” the team noted.
The final data set, dubbed RobustPDx, included information from 261 participants: 73 with self-reported Parkinson’s, 33 who self-reported suspected Parkinson’s, and 155 controls who did not have Parkinson’s. In addition to the data gathered from the online platform, the researchers also collected info on participants’ demographics, as well as the type of computer they were using and their handedness.
“As remote digital health assessments increasingly rely on AI, it is important to evaluate whether these models remain robust across real-world sources of variation, including differences in device type and handedness,” the researchers noted.
The scientists are hopeful that this detailed data set will be useful for future studies aimed at developing AI tools for remote diagnosis and monitoring of Parkinson’s. The researchers wrote that their data set “supports the study of robustness and reliability challenges that may arise when digital health models are deployed across [variable] users, devices, and interaction patterns.”
“Thus, this dataset is intended not only for measuring overall classification performance but also as a benchmark for investigating the robustness and generalizability of AI models in [varied] remote digital health settings,” they added.
Leave a comment
Fill in the required fields to post. Your email address will not be published.