Algorithms Detect Parkinson’s Based on Verily Study Watch Data
Study combines machine learning with sensor data from the PPMI
A new analysis using machine-learning and sensor data from the Parkinson’s Progression Markers Initiative (PPMI), a digital health research program sponsored by The Michael J. Fox Foundation for Parkinson’s Research (MJFF), successfully distinguished between people with and without Parkinson’s disease, according to a small study led by Cohen Veterans Bioscience (CVB).
Specific information, such as motion data, was recorded using the Verily Study Watch, a wrist-worn device used by the participants for up to 23 hours a day for several months.
“This study shows the feasibility of leveraging unconstrained and unlabeled wearable sensor data to accurately detect Parkinson’s disease using powerful deep learning methods,” Lee Lancashire, PhD, principal investigator of the study and chief information officer at CVB, said in a press release.
“Through this combination of wearables and [artificial intelligence], we are one step closer to monitoring individual healthcare-related activity, such as motor function outside the clinic, unleashing the potential for early detection and diagnosis of diseases such as Parkinson’s disease,” he added.
The study, “Deep Learning for Daily Monitoring of Parkinson’s Disease Outside the Clinic Using Wearable Sensors,” was published in the journal Sensors.
Parkinson’s disease is a progressive neurological condition characterized by motor symptoms, such as bradykinesia (slow movements), walking abnormalities, and tremors. Diagnosis can be difficult especially because there are no objective biomarkers for the disease. Better methods of tracking disease progression also are needed so that clinicians can deliver individualized care and treatments.
According to researchers, sensor technology is quickly developing. Many studies have focused on looking for digital biomarkers linked with specific movement features of Parkinson’s. However, data from these studies were recorded under controlled lab settings and do not reflect patients’ movements in their daily environment.
Now, a team of researchers sought to collect data with the Verily Study Watch, worn daily by a subset of participants in the PPMI study, and determine whether their newly developed computer-generated algorithms can be used to identify people with Parkinson’s disease based on walk-like events.
The PPMI (NCT01141023) is a longitudinal, observational study of people with and without Parkinson’s disease. Its goal is to identify biomarkers associated with Parkinson’s risk, onset, and progression.
In 2018 the PPMI launched a substudy using the Verily Study Watch at sites in the U.S.; all subjects enrolled in PPMI were invited to participate.
“Thus, compared to other data types associated with the full PPMI dataset, the wearable data are not restricted to following the progression of [Parkinson’s] starting at an early, untreated stage but may begin at any point along the trajectory,” the researchers wrote.
For the new analysis, researchers extracted participant data from 11 people from the PPMI database on June 2021: seven clinically diagnosed patients with Parkinson’s and four controls. Among those with Parkinson’s, five had genetic risk variants in the LRKK2, GBA, and SNCA genes, and two were recently diagnosed patients who remained untreated during the study.
Patients were asked to wear the Verily Study Watch for up to 23 hours a day from several months to two years during their daily activities.
The new algorithm showed an accuracy of 100% for a Parkinson’s diagnosis based only on the data from participants’ walking movements accumulated over a day. According to the researchers, this can be interpreted “as the ability to identify subtle gait changes related to [Parkinson’s] that are unaccounted for in the UPDRS [Unified Parkinson’s Disease Rating Scale] scores.” The UPDRS is a scale commonly used to rate the severity of symptoms in Parkinson’s disease.
It also could distinguish with near 90% accuracy between people with and without a Parkinson’s diagnosis on single, five-second, walk-like movements.
“Although additional studies are needed, we are excited about the potential of using sensor data obtained through a patients’ normal activity to enable physicians to monitor and classify [Parkinson’s] symptoms through easy to obtain objective measures which can be used to improve clinical decision making and guide therapeutic interventions,” said Mark Frasier, PhD, co-author of the study, the chief scientific officer at MJFF.
Funds from CVB and a grant from MJFF supported this study.