Distinguishing Parkinsonian Rest Tremor from Voluntary Hand Movements through Subthalamic and Cortical Activity
Parkinson’s disease (PD) is a common neurodegenerative disorder characterized mainly by resting tremor, bradykinesia, and rigidity. Deep Brain Stimulation (DBS) has been widely used to treat the motor symptoms of PD (Krauss et al., 2021). However, DBS treatment also has significant side effects, most of which are caused by the extension of stimulation to areas surrounding the DBS target structures (Koeglsperger et al., 2019). To reduce these side effects, researchers have proposed an adaptive Deep Brain Stimulation (aDBS) scheme that adjusts the intensity and timing of DBS by real-time monitoring of the patient’s current motor state (Little et al., 2016; Piña-Fuentes et al., 2017; Tinkhauser et al., 2017). Particularly for the treatment of resting tremors in PD, aDBS holds the promise of significantly improving therapeutic outcomes while reducing side effects.
In traditional DBS control signals, research mainly focuses on β-band activity in the subthalamic nucleus (Gilron et al., 2021). However, with further research, scientists have realized that tremors are not solely determined by the activity of the subthalamic nucleus but involve broader cortico-pathway networks. Therefore, studying the brain signals of PD tremors may reveal more valuable information to improve the efficacy of aDBS. This study aims to distinguish between resting tremor and voluntary hand movements in PD patients by analyzing the electroencephalographic activity under various motor states.
Source of Study
This paper was authored by Dmitrii Todorov, Alfons Schnitzler, Jan Hirschmann, and others, affiliated with the Institute of Clinical Neuroscience and Medical Psychology and the Department of Neurology at Heinrich Heine University in Düsseldorf, Germany, the Centre de Recherche en Neurosciences de Lyon - Inserm U1028 in France, and the Centre de Recerca Matemàtica in Spain. The research findings were published in the Clinical Neurophysiology journal (accepted on October 31, 2023, issue no. 157 for the year 2024).
Research Process
Subjects and Data Acquisition
This study re-analyzed previously collected data derived from magnetoencephalography (MEG) and subthalamic nucleus (STN) local field potential (LFP) recordings of six PD patients. These patients underwent DBS implantation surgery a day before testing, and all had a diagnosis of idiopathic PD. The study was conducted under both medication off (med off) and medication on (med on) conditions for each patient to evaluate the differences in brain activity under these states.
Data Recording
Data was recorded on the second day post-surgery and included four segments: rest, Task 1, rest, and Task 2. Task 1 involved forearm extension, and Task 2 involved self-paced fist-clenching. Each segment was recorded for 5 minutes, and all movements were performed by the body side with significant symptoms. Rest and task modes experienced natural emergence and disappearance of tremors. The recorded data included LFP from the subthalamic nucleus, MEG, and electromyography (EMG) from the forearm with a sampling rate of 2000 Hz.
Data Processing and Feature Extraction
For efficient data processing, MEG and LFP data were resampled to 256 Hz. The research team applied several preprocessing steps to ensure data quality, including MEG artifact detection, spatiotemporal signal space separation, and LFP artifact detection. For feature extraction, the team used a non-overlapping 1-second time window to compute short-term variance, known as Hjorth activity. This method’s advantage is avoiding overfitting while facilitating comparison of feature information from different brain regions.
Machine Learning Model
To distinguish between the four different motor states (rest, tremor, fist-clenching, and forearm extension), the research team applied the XGBoost algorithm based on gradient-boosted tree learning and used 5-fold cross-validation for classification. To balance the amount of data in different states, they employed balanced accuracy as the performance metric and used oversampling to address data imbalance.
Electrode Localization and Hyperparameter Tuning
The research team used the Lead-DBS software package to reconstruct the DBS electrode placement for accurate localization. To enhance model performance, they tuned the hyperparameters of the XGBoost algorithm, including selecting the optimal LFP channels and adjusting model parameters.
Main Results
Single-Nucleus Data Analysis
Classification performance based solely on subthalamic nucleus LFP data was weak, correctly classifying rest at 62%, while the correct classification rates for tremor, fist-clenching, and forearm extension were 32%, 34%, and 23%, respectively. These poor results indicate that subthalamic nucleus activity alone is insufficient to accurately distinguish motor states.
Nucleus-Cortex Joint Analysis
When combining subthalamic nucleus and cortical activity features for classification, model performance improved significantly, with correct classification rates of 74% for rest, 58% for tremor, 80% for fist-clenching, and 81% for forearm extension. This indicates that the inclusion of cortical signals greatly enhances classification performance, with sensory-motor cortex data being particularly crucial.
Effect of Dopaminergic State on Classification Performance
The results also indicated that different dopaminergic states had little impact on classification performance, providing solid theoretical support for practical applications.
Conclusion and Significance
This study reveals that the combination of subthalamic nucleus and cortical signals can significantly distinguish between resting tremor and voluntary hand movements in PD, providing strong support for developing more accurate aDBS systems. Specifically:
Scientific Value: A more comprehensive understanding of the neuro mechanisms of tremor and voluntary movement may further optimize PD treatment protocols, reducing unnecessary DBS stimulation and thereby minimizing side effects.
Application Value: The study results lay the foundation for developing more intelligent DBS systems, such as real-time monitoring systems that integrate recordings from the cortex and subthalamic nucleus.
Innovative Points: This study not only combines cortical and subthalamic nucleus signals but also utilizes advanced machine learning algorithms for classification, significantly enhancing the accuracy in distinguishing different motor states of PD.
Highlights
- Significant Improvement in Classification Performance: Introducing cortical signals greatly enhances classification performance, especially in the sensory-motor cortex region.
- Prospects for Adaptive DBS: Combining cortical and subthalamic nucleus signals with machine learning for real-time motor state monitoring shows promise for developing more precise aDBS systems.
- Minor Impact of Medication State: The study indicates that different dopaminergic medication states have minimal impact on classification performance, enhancing the practical feasibility of the method in clinical applications.
Other Valuable Information
The study also shows that incorporating signals from broader brain regions can further improve classification performance, despite the significant contribution of the sensory-motor cortex region. Future research can further optimize signal extraction and machine learning models to improve real-time classification accuracy and robustness. Additionally, practical applications could explore integrating these research findings into clinical environments, particularly using novel implantable DBS systems for real-time brain signal monitoring and stimulation adjustment.
This study provides scientists with new ideas and methods for developing more intelligent treatment solutions for PD. Future research combining larger sample sizes and advanced algorithms may lead to greater breakthroughs in early diagnosis and personalized treatment.