AI separates Parkinson’s brain rhythms from background activity, capturing rapid changes

An AI method has picked out fast changes in Parkinson’s-related brain activity that conventional analysis can miss. Researchers report that NONOS separates repeating brain rhythms from background activity without blurring their timing. In a NeuroImage study, that separation improved prediction of patients’ next step and revealed rapid responses to electrical stimulation in mice.

The work was led by Ji-Woong Choi at the Daegu Gyeongbuk Institute of Science and Technology, or DGIST. Collaborators included Kumoh National Institute of Technology, Korea University and the Korea Brain Research Institute. DGIST described the findings in a November 4 account published by Medical Xpress, which identifies the paper as “Neural oscillation and non-oscillation separator for highly time-resolved decomposition of beta band activity.”

Two ways conventional analysis can lose the signal

Brain electrical recordings mix oscillations—patterns that repeat at regular intervals—with activity that does not repeat in the same way. The distinction matters in Parkinson’s disease because unusually strong beta rhythms are associated with slowed movement. These waves repeat dozens of times a second, and their strength is used to assess treatment effectiveness.

Common filtering techniques measure beta-wave strength without accounting for the background. That creates a risk of interpreting a background shift as a change in the rhythm itself. Newer approaches remove background activity, but group recordings into fixed time windows. Those windows can obscure brief changes, trading a cleaner separation for a less precise view of when something happened.

NONOS takes a different route. Its deep-learning model receives a brain recording over time and extracts background components using their characteristic relationship between signal power and frequency. The researchers then subtract that estimated background from the original recording to isolate the rhythms. They say this preserves temporal resolution: the ability to distinguish changes occurring close together in time.

A next-step prediction and a mouse stimulation response

The human analysis tested whether the extracted rhythms helped predict which foot a person with Parkinson’s would move forward. The team reported a prediction score of 0.98, compared with 0.54–0.79 for conventional methods. The institutional account describes values closer to one as better performance; it does not identify the metric by name, so the figure should not be read as a stated percentage accuracy.

0.98NONOS-based signals

The researchers’ reported prediction score using rhythms extracted by NONOS.

0.54–0.79Conventional methods

The comparison scores reported by the team; values closer to one indicate better performance.

A separate experiment within the same study examined mice modeling Parkinson’s disease. After deep brain stimulation—therapy that applies electrical stimulation—the researchers observed a rapid drop in beta-wave strength. They report that NONOS was the only method in their comparison sensitive enough to detect that fast decrease, which conventional methods had difficulty identifying.

The implantable version still needs development

Choi’s proposed next step connects the signal-analysis result to personalized treatment. He says reducing NONOS’s computational requirements could allow it to run in real time on implantable brain devices. Such a system would aim to improve stimulation that adjusts to each patient’s condition. Real-time operation inside an implant is a future goal, not a capability demonstrated in this study.