Toward a Precision Medicine Approach to Deep Brain Stimulation
Key findings
- Applying machine learning to analyzing intracranial signals might revolutionize adaptive deep brain stimulation by personalizing therapy to patients' specific symptoms and behaviors
- This study examined patient-specific, machine learning–based models for analyzing movement during invasive neurophysiology based on both the frequency and spatial dimensions of recorded brain activity in 11 patients with Parkinson's disease
- Adding spatial information extracted from multichannel recordings to machine learning models showed potential to improve movement decoding compared to a single-channel approach
- Some brain maps were similar across patients, but each patient had their own spatial map at each frequency band
- For the development of clinically useful brain–computer interfaces, individualized assessment of features and model performance will be required
Continuous high-frequency deep brain stimulation (DBS) improves motor symptoms of Parkinson's disease, but side effects such as dysarthria and dyskinesia can complicate treatment. Closed-loop DBS, also called adaptive DBS, is guided by the patient's individual electroclinical state and might minimize side effects by reducing the amount of unnecessary stimulation delivered to the brain.
Subscribe to the latest updates from Neuroscience Advances in Motion
Adaptive DBS devices are invasive brain–computer interfaces in which computer algorithms informed by brain signals adapt the amount of stimulation. A current effort in the field is to augment adaptive DBS with machine learning—for example, use it for movement decoding (analyzing movement-related oscillations on recordings)—thereby refining brain stimulation.
Now, researchers at Massachusetts General Hospital have shown movement decoding could be improved by the use of spatial information and multimodal brain recordings. Victoria Peterson, PhD, of the Department of Neurosurgery, Mark Richardson, MD, PhD, director of the Functional Neurosurgery Program, and colleagues explain in Experimental Neurology.
Background
Most models for movement decoding in the context of adaptive DBS are based on single-channel frequency domain features. That strategy disregards the spatial information available in multichannel recordings. This study examined whether considering both multichannel and multiple-site recordings could better correlate the neural source with a movement task.
Methods
The dataset used in the study was described in a previously published Brain paper. It comprised subthalamic nucleus local field potentials (STN-LFPs) and subdural electrocorticography recordings simultaneously acquired from 11 patients with Parkinson's disease as they underwent awake DBS implantation. Patients were asked to press a hand-grip force transducer with either their right or left hand after a visual cue appeared.
The researchers constructed patient-specific, supervised machine learning–based neural decoding models that were based on both the frequency and spatial dimensions of the recorded brain activity.
They compared this spatio-spectral approach to a single-channel approach. They also compared decoding models trained on the combination of electrocorticographic and STN-LFP recordings with models trained only on electrocorticography.
Results
The principal findings were that:
- Intracranial spatial information can be added to models from improving movement decoding
- Machine learning models based on multichannel recordings showed potential to improve movement decoding compared to a single-channel approach
- Electrocorticography was superior to STN-LFPs for movement decoding, and there was no gain when adding STN-LFP features to electrocorticographic features
- Although some brain maps were similar across patients, each patient had their own spatial map at each frequency band
Guidance for Future Development Efforts
Cross-patient movement decoding would be useful for "out-of-the-box" use of machine learning–based brain–computer interfaces. However, this study emphasizes that spatio-spectral patterns are individualized, which will make cross-patient implementation difficult.
Combining two recording modalities will improve both the decoding capacity of the model and model robustness in some cases. Still, individual assessment of features and model performance will be required for optimal decoding.
view original journal article Subscription may be required
Visit the Functional Neurosurgery Program
Refer a patient to the Department of Neurosurgery