When measuring progressive motor impairments, namely ataxia, patients and clinicians need regular and objective monitoring. However, the current gold standard of clinician-based assessment is neither regular nor completely objective. Even previous attempts to have technology capture movement irregularities are susceptible to having human errors baked in. Now, a UMass Amherst-led team, with Harvard Medical School/Mass General Brigham and the University of Alabama at Birmingham, has developed a new artificial intelligence (AI) model that cuts out the subjective noise for more accurate ataxia assessment that could one day be incorporated into wearable devices.

Cerebral ataxia is a neurological impairment in the cerebellum that results in a wide range of irregular or uncoordinated movement, including walking and balance issues, muscle weakness, over or undershooting the indented position of the moving body part, and speech or swallowing impairment. It can stem from genetic conditions, damage to the cerebellum (which can be caused by stroke, brain tumor, multiple sclerosis or degenerative diseases), alcohol misuse and certain prescription drugs.

Image

“Most ataxias are progressive, so patients are getting worse over time, so it’s important to frequently measure their severity and how their diseases progress,” says Juhyeon Lee, first author on the paper and computer science Ph.D. candidate in the Advanced Human Health Analytics (AHHA) Laboratory. However, the only way to currently measure ataxia is through in-person hospital visits, often with a specialized neurologist.

Wearable devices are poised to be an unobtrusive tool for continuous motor function monitoring and an accessible alternative to clinician-based assessment. Yet this technology is still limited in its capability to convert movement signals into consistent movement scores. To do so requires data processing by AI. 

Previous attempts to generate a motor severity score from movement data has trained AI using clinician-evaluated motor assessments as the gold standard. The model is fed the relevant features (which are either manually engineered or machine learned) which it then uses to evaluate movement data. When tested against a known data set of ataxia movement scores, if the AI model successfully generates a score that matches the clinician’s score, it is deemed successful. 

While clinician assessments should be consistent between patients, any time a human is asked to make a judgment, there is always a chance for variability. And an AI model is only as good as the data used to train it. A model that uses clinician assessment as its gold standard runs the risk of having their subjectivity baked into its own outputs.

To avoid this bias, Lee used a contrastive learning model to capture implicit differences between impaired or healthy motor function. The AI initially uses the clinician scores to form similar (positive) and different (negative) data pairs. Then the AI is further trained to compare the sensor data, independent of the exact clinician’s score, to differentiate between positive and negative pairs. 

Even if there are measurement errors in clinician-evaluated reference scores, this contrastive model can still identify precise and sensitive patterns to generate a score that better reflects reality, rather than replicating clinical scores.

“Consider two patients with the same underlying motor severity but slightly different clinical scores due to measurement errors,” the researchers write in their paper. A contrastive model can learn to recognize the true similarity of the performance as captured by the sensor data, ultimately producing ataxia scores that may better reflect the reality of the underlying motor impairment.

The researchers collected data from 87 people with ataxias and 44 neurologically healthy individuals and completed a finger-to-nose task (a common test for clinicians to evaluate ataxia severity) while wearing wrist sensors to capture movement. A neurologist also evaluated the participants using a Brief Ataxia Rating Scale (BARS). The new model achieved high correlations with clinician scores, good sensitivity to changes over time and strong reliability within sessions.

Image

The researchers envision that this machine learning framework could be applied broadly by adapting it to other tasks, body parts, or even sensor inputs, such as speech data from microphones. 

“For individuals living with progressive neurological conditions like ataxia, not knowing how their disease progresses over time can be incredibly frustrating,” says Ivan Lee, associate professor in the Manning College of Information and Computer Sciences at UMass Amherst, and principal investigator of the AHHA Lab. “Our work demonstrates that AI-driven wearables can bridge this gap. By providing a scalable, objective measure of motor impairment, we can equip patients, their families, and their clinical teams with the continuous data they need to collaboratively manage their care.”

This research, conducted in collaboration with Anoopum S. Gupta, assistant professor of neurology at Harvard Medical School and neurologist at Mass General Brigham, and Brandon Oubre, assistant professor of computer science at University of Alabama at Birmingham, is presented in IEEE Transactions on Biomedical Engineering. The paper was recognized with a Featured Article Award for its novelty in applying this contrastive learning model over hand-engineered data to generate ataxia severity scores from wearable sensors.