📄 bioengineering

Device-embedded accelerometry complements neural signals for tracking parkinsonian motor states

This study demonstrates that device-embedded accelerometry provides a more robust and accurate biomarker for tracking Parkinsonian motor states across varying stimulation conditions compared to neural signals, supporting its integration into next-generation adaptive deep brain stimulation systems.

Original authors: LIU, T., Yao, J., Abdi-Sargezeh, B., Sharma, A., Lasbareilles, C., Tsi Lok Ho, R., Cheung, J., Denison, T., Tan, H., Neumann, W.-J., Zhu, M. M., Liu, S., Starr, P., Little, S., Oswal, A.

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: LIU, T., Yao, J., Abdi-Sargezeh, B., Sharma, A., Lasbareilles, C., Tsi Lok Ho, R., Cheung, J., Denison, T., Tan, H., Neumann, W.-J., Zhu, M. M., Liu, S., Starr, P., Little, S., Oswal, A.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Picture: Fixing the "Smart" Parkinson's Treatment

Imagine Deep Brain Stimulation (DBS) as a high-tech thermostat for the brain. For people with Parkinson's disease, this "thermostat" sends electrical pulses to calm down a specific part of the brain (the Subthalamic Nucleus, or STN) that is misfiring and causing tremors, stiffness, and slow movement.

Currently, most of these thermostats run on a timer—they just keep humming along. The goal of "Adaptive DBS" is to make them smart: they should turn up the power when the patient is stiff and turn it down when they are moving well. To do this, the machine needs a "thermometer" to read the brain's temperature (the symptoms).

The Problem: The traditional thermometer is a wire inside the brain listening to electrical signals. But here's the catch: when the machine turns on the electricity to fix the problem, it creates so much "static noise" that the thermometer can't hear the brain clearly anymore. It's like trying to listen to a whisper while someone is shouting right next to your ear.

The Solution: This paper suggests adding a second thermometer: a tiny motion sensor (accelerometer) already built inside the device. The researchers found that this motion sensor is actually better at telling the machine how the patient is feeling, even while the electricity is blasting.

The Experiment: A 1,900-Hour Listening Party

The researchers studied 11 people with Parkinson's who had special implants that could both send electricity and record data. They recorded over 1,900 hours of data while the patients went about their normal daily lives (eating, walking, talking).

They had three things running at the same time:

Brain Wires: Listening to the STN and the motor cortex (the brain's movement center).
The Built-in Motion Sensor: The accelerometer inside the chest implant, feeling the patient's body move.
Wrist Watches: Special watches (called PKG) worn by the patients to objectively measure how stiff or shaky they were.

They looked at two main problems:

Bradykinesia: Being too slow and stiff.
Dyskinesia: Moving too much involuntarily (like uncontrollable dancing).

Key Finding 1: The "Total Noise" vs. The "Specific Rhythm"

For years, doctors have tried to use the total amount of "beta waves" (a specific brain rhythm) in the STN as the signal for stiffness.

The Analogy: Imagine a crowded room. The "total beta power" is like measuring the total volume of the room.
The Discovery: The researchers found that the total volume is a bad indicator because it mixes two different things:
1. The Rhythm (Periodic): A steady drumbeat that gets louder when the patient is stiff.
2. The Hiss (Aperiodic): A background static hiss that actually gets quieter when the patient is stiff.
The Result: Because these two things move in opposite directions, adding them together (the "total volume") cancels each other out. It's like trying to measure the temperature by mixing boiling water and ice water; the result tells you nothing.
The Fix: When they separated the "rhythm" from the "hiss," they found the rhythm was a good sign of stiffness, but the "hiss" (aperiodic activity) was actually a very stable signal that didn't get confused by the treatment.

Key Finding 2: The Motion Sensor is the MVP

The most surprising result was about the accelerometer (the motion sensor inside the chest implant).

The Analogy: Think of the brain wires as a seismograph trying to guess if an earthquake is happening by listening to the ground shake. Think of the accelerometer as a smartwatch that actually feels the shaking.
The Discovery:
- When the brain was not being stimulated, the brain wires were okay at guessing the symptoms.
- But when the stimulation was turned ON (which is when the machine needs to make decisions), the brain wires got confused by the electrical noise. Their ability to predict symptoms dropped significantly.
- The accelerometer, however, didn't care about the electrical noise. It kept tracking the patient's stiffness and shaking perfectly, whether the machine was on or off.
The Verdict: In head-to-head tests, the motion sensor was better at predicting symptoms than the brain wires, especially when the treatment was active.

Key Finding 3: The Best Team is a Hybrid

The researchers tried to build a computer model to predict symptoms using different combinations of data.

Brain Only: Okay, but gets confused when the machine is on.
Motion Sensor Only: Very good, very stable.
Brain + Motion Sensor: The best combination.

The Metaphor: Imagine you are trying to navigate a foggy road.

The Brain Wires are like a GPS that sometimes loses signal in the fog (when the stimulation is on).
The Motion Sensor is like a driver looking out the window; they can always see the road.
The Solution: Use the driver (motion sensor) to steer the car, but keep the GPS (brain wires) as a backup to give extra details when the road is clear.

Why This Matters (According to the Paper)

The paper concludes that we shouldn't rely only on brain waves to control these smart implants because the treatment itself messes up the brain waves.

Instead, the device should use the built-in motion sensor as its primary "eyes" to see how the patient is moving. This sensor is robust, doesn't get confused by the electricity, and is already inside the device, so no new surgery or external hardware is needed.

In short: The paper proves that a tiny motion sensor inside the implant is a superior, reliable way to tell a Parkinson's treatment machine when to turn up or down the power, even while the machine is actively working.

Technical Summary: Device-embedded Accelerometry Complements Neural Signals for Tracking Parkinsonian Motor States

Problem Statement
Adaptive Deep Brain Stimulation (aDBS) for Parkinson's disease (PD) relies on physiological biomarkers to infer motor state and guide therapeutic stimulation. However, a critical limitation exists: the neural biomarkers used for control (such as subthalamic nucleus [STN] beta power) are themselves altered by the stimulation they are intended to guide. This "self-interference" can degrade the utility of neural signals for closed-loop control. Furthermore, standard clinical assessments lack the temporal resolution to capture motor state fluctuations at the timescale of neural dynamics. While wearable accelerometers (e.g., Parkinson's KinetiGraph) have shown promise, they introduce challenges regarding user compliance and real-time signal integration with implantable pulse generators (IPGs). The study addresses whether accelerometers already embedded within modern DBS devices can reliably track bradykinesia and dyskinesia severity across different stimulation conditions, potentially offering a robust behavioral biomarker that complements or outperforms neural signals.

Methodology
The study analyzed over 1,900 hours of chronic, simultaneous recordings from 11 patients with idiopathic Parkinson's disease implanted with an investigational sensing-enabled DBS system (Medtronic Summit RC+S).

Data Acquisition: Recordings included local field potentials (LFPs) from the STN and sensorimotor cortex, device-embedded triaxial accelerometry from the IPG, and concurrent wearable motor scores from bilaterally worn Parkinson's KinetiGraph (PKG) devices. Data were collected during activities of daily living in both OFF-stimulation (986 hours) and continuous ON-stimulation (915 hours) states.
Signal Processing: Neural signals were processed to extract spectral features, separating periodic (oscillatory) and aperiodic (1/f) components using the FOOOF algorithm. Features included band-specific power (beta, gamma), aperiodic offset/exponent, and cortico-subthalamic coherence. Accelerometer data were processed to extract 30 features, including entropy, jerk, and Euclidean norm, aligned with 2-minute PKG intervals.
Analysis: The authors employed correlation analyses (Spearman, Pearson) and four regression models (Random Forest, SVM, ElasticNet, Fully Connected Neural Network) to predict PKG-derived bradykinesia and dyskinesia scores. Model performance was evaluated using stratified 5-fold cross-validation ( $R^2$ ). Statistical significance was assessed using Benjamini–Hochberg False Discovery Rate correction.

Key Contributions

Decomposition of Neural Biomarkers: The study demonstrates that "total" STN beta power is a less reliable biomarker because it conflates opposing physiological processes: periodic beta activity (which correlates with worsening bradykinesia) and aperiodic activity (which correlates with improvement).
Validation of Device-Embedded Accelerometry: The authors provide the first demonstration that IPG-embedded accelerometers can robustly track motor symptom fluctuations (both bradykinesia and dyskinesia) with high sensitivity and specificity, even during continuous stimulation.
Differential Robustness: The study identifies that while neural biomarkers (particularly periodic beta) suffer from decoupling from motor state under active stimulation, device-embedded accelerometry and specific neural features (STN aperiodic parameters, cortical periodic power, and cortico-subthalamic coherence) remain comparatively stable.
Superior Predictive Performance: Machine learning models utilizing device-embedded accelerometry features outperformed models using only neural features for decoding motor states, both in OFF and ON stimulation conditions.

Results

Neural Dynamics: Off-stimulation, periodic STN beta power increased with worsening bradykinesia, while aperiodic power decreased. Under continuous stimulation, the correlation between periodic STN beta power and symptom severity was significantly attenuated. Conversely, STN aperiodic parameters, cortical periodic activity, and cortico-subthalamic coherence maintained stable relationships with motor symptoms during stimulation.
Accelerometry Performance: Specific accelerometer features (entropy for bradykinesia; jerk for dyskinesia) showed strong correlations with PKG scores. In regression tasks, accelerometer-only models achieved higher prediction accuracy ( $R^2 \approx 0.24$ for bradykinesia, $0.31$ for dyskinesia) than neural-only models ( $R^2 \approx 0.11-0.22$ ).
Combined Models: Integrating neural and accelerometry features yielded the highest performance ( $R^2 \approx 0.29$ for bradykinesia, $0.35$ for dyskinesia). Notably, a combination of STN signals and accelerometry approached the performance of the full multimodal model, suggesting a practical configuration for clinical systems lacking cortical recordings.
Stimulation Effects: While neural model performance declined modestly under stimulation, accelerometry-based models maintained their predictive power, effectively compensating for the degradation of neural biomarkers.

Significance and Claims
The paper claims that device-embedded accelerometry represents a robust, stimulation-resilient behavioral biomarker for motor state tracking in Parkinson's disease. By demonstrating that accelerometry outperforms neural features in decoding symptom severity—particularly when stimulation alters neural dynamics—the study motivates the integration of behavioral readouts into next-generation adaptive DBS systems. The authors suggest a multimodal approach where accelerometry provides a stable estimate of clinical state, complementing neural biomarkers that offer mechanistic specificity. This approach could address the limitations of current aDBS systems where control signals are compromised by the very therapy they regulate. The findings support the use of aperiodic neural components and network-level coherence as more stable neural targets than total beta power, while highlighting the practical viability of using existing IPG hardware for closed-loop control without requiring external wearables.