AI detects hidden movement clues linked to Parkinson’s disease

Scientists have developed an artificial intelligence system capable of identifying subtle signs of brain disorders, including Parkinson’s disease, by analyzing simple video recordings of a person’s hand movements. The technology successfully pinpointed minute motor impairments in individuals whose performance had been judged as perfectly normal by expert neurologists, opening a new avenue for detecting neurodegenerative conditions at their earliest stages. The findings are published in Nature.

Researchers are persistently searching for better ways to identify neurodegenerative diseases like Parkinson’s disease before obvious symptoms appear. Early detection can be essential for managing the condition and developing treatments that might slow its progression. A major challenge is that the initial changes in motor function are often too slight for a doctor to see during a standard clinical examination. This makes it difficult to diagnose the disease or to identify individuals who are at high risk.

One such high-risk group includes people with a condition known as idiopathic REM Sleep Behavior Disorder. This is a sleep disorder where individuals physically act out their dreams, sometimes with vigorous movements. A large majority of people diagnosed with this sleep disorder will eventually develop Parkinson’s disease or a related condition. This makes them an ideal population for studying the earliest, or prodromal, signs of neurodegeneration.

The core motor symptom of Parkinson’s disease is known as bradykinesia, which is a general term for slowness of movement. This can be accompanied by other related issues, such as hypokinesia, which refers to movements that are too small in amplitude. Another key indicator is the “sequence effect,” where repetitive movements become progressively slower or smaller over time.

While these features are hallmarks of established Parkinson’s disease, it has been unclear if they are present in a detectable way in very early stages or in at-risk individuals with REM Sleep Behavior Disorder. Current methods for quantifying these subtle changes often require specialized, expensive equipment like accelerometers or optical trackers, limiting their widespread use.

To overcome these barriers, a team of researchers from the University of Florida sought to determine if an automated analysis of standard video recordings could unmask these hidden motor deficits. They hypothesized that their video-based approach could detect movement problems in people with early Parkinson’s disease even when a clinician could not, and that some of these features, especially the sequence effect, would also be present in individuals with the sleep disorder.

To test their ideas, the researchers recruited a total of 66 participants, who were divided into three groups. The first group consisted of 18 patients who had been diagnosed with early-stage Parkinson’s disease. The second group was made up of 16 individuals who had a confirmed diagnosis of idiopathic REM Sleep Behavior Disorder. The final group included 32 healthy adults with no history of neurological or sleep disorders to serve as a control group.

All participants underwent a standard neurological motor examination, which was recorded on video using a consumer-grade camera. A central part of this exam is the Finger Tapping task, where a person is asked to repeatedly tap their index finger against their thumb as quickly and as widely as possible for about ten seconds.

A fellowship-trained movement disorders neurologist evaluated and scored each participant’s performance on the Finger Tapping task using a standard clinical scale from 0 to 4, where a score of 0 signifies normal movement and a score of 4 indicates severe impairment. For their analysis, the researchers selected only the videos of finger-tapping performances that received a perfect score of 0. This selection ensured that any motor impairments the artificial intelligence might find would be truly subclinical, meaning they were invisible to the trained human eye.

The selected videos were then processed using a specialized software tool that employs a deep learning model. This model automatically identified and tracked the positions of 21 distinct points on the hand in every frame of the video. From this tracking data, the system calculated the distance between the tip of the index finger and the tip of the thumb throughout the task. This measurement created a continuous signal that represented the opening and closing of the fingers over time.

The researchers then extracted four key kinematic features from this signal for each participant: average movement amplitude (how wide the fingers opened), average movement speed, the decrement in amplitude (how much the movement size decreased from the beginning to the end of the task), and the decrement in speed (how much the movement speed declined).

The artificial intelligence system revealed clear differences between the groups that were not apparent from the clinical scores. Individuals with Parkinson’s disease showed significantly smaller movement amplitude and slower movement speed compared to both the healthy controls and the individuals with REM Sleep Behavior Disorder. They also exhibited a pronounced sequence effect, with both their movement size and speed decreasing over the course of the repetitive taps.

The results for the group with REM Sleep Behavior Disorder were particularly revealing. Unlike the Parkinson’s group, their average movement amplitude and speed were not different from the healthy control group. However, they did show a significant decrement in both amplitude and speed, a pattern similar to what was seen in the Parkinson’s patients. This suggests that the sequence effect, the fatigue-like decline in performance during repetitive movements, may be one of the earliest motor signs of the underlying disease process, appearing even before the classic slowness and smallness of movement associated with a Parkinson’s diagnosis.

To further test the power of these hidden measurements, the researchers used a machine learning algorithm known as a random forest to see if it could accurately classify individuals into their respective groups based only on the four video-derived features. The algorithm performed with high accuracy. It could distinguish people with Parkinson’s disease from healthy controls with 81.5% accuracy.

It could also differentiate individuals with REM Sleep Behavior Disorder from healthy controls with 79.8% accuracy. When tasked with separating the two clinical groups, the model could tell apart individuals with the sleep disorder from those with Parkinson’s disease with 81.7% accuracy. These classification results demonstrate that the subtle, computer-detected features contain enough information to reliably separate the groups, even when they appear identical to a human expert.

The researchers acknowledge certain limitations of their study, most notably the relatively small number of participants. Findings from a smaller sample need to be validated in larger, more diverse groups of people to ensure they are generalizable. Additionally, this work focused exclusively on motor symptoms derived from video. Future research could aim to combine this accessible video analysis with other biological markers, such as those from brain imaging or cerebrospinal fluid, to create an even more powerful and comprehensive tool for predicting who is at risk of developing Parkinson’s disease.

Nevertheless, this study provides evidence that automated video analysis can serve as a sensitive, low-cost, and accessible tool for detecting the earliest signs of neurodegeneration. Such a technology could one day be used for large-scale screening to identify at-risk individuals for inclusion in clinical trials for new neuroprotective therapies, well before more pronounced and life-altering symptoms begin to manifest.

The study, “Video analysis reveals early signs of Bradykinesia in REM sleep behavior disorder and Parkinson’s disease,” was authored by Diego L. Guarín, Gabriela Acevedo, Carolina Calonge, Joshua K. Wong, Nikolaus R. McFarland, Adolfo Ramirez-Zamora, and David E. Vaillancourt.