Concurrent decoding of distinct neurophysiological fingerprints of tremor and bradykinesia in Parkinson’s disease
Curation statements for this article:-
Curated by eLife
eLife assessment
This important study aimed to identify rhythms linked to primary symptoms of Parkinson's disease such as involuntary shaking of the limbs and slowness. The evidence supporting the conclusions is solid, although validating their behavioural measures and considering the relationship between signatures from different brain regions would have strengthened the study. The work will be of broad interest to movement control, movement disorders, and brain stimulation fields.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
Abstract
Parkinson’s disease (PD) is characterized by distinct motor phenomena that are expressed asynchronously. Understanding the neurophysiological correlates of these motor states could facilitate monitoring of disease progression and allow improved assessments of therapeutic efficacy, as well as enable optimal closed-loop neuromodulation. We examined neural activity in the basal ganglia and cortex of 31 subjects with PD during a quantitative motor task to decode tremor and bradykinesia – two cardinal motor signs of PD – and relatively asymptomatic periods of behavior. Support vector regression analysis of microelectrode and electrocorticography recordings revealed that tremor and bradykinesia had nearly opposite neural signatures, while effective motor control displayed unique, differentiating features. The neurophysiological signatures of these motor states depended on the signal type and location. Cortical decoding generally outperformed subcortical decoding. Within the subthalamic nucleus (STN), tremor and bradykinesia were better decoded from distinct subregions. These results demonstrate how to leverage neurophysiology to more precisely treat PD.
Article activity feed
-
-
eLife assessment
This important study aimed to identify rhythms linked to primary symptoms of Parkinson's disease such as involuntary shaking of the limbs and slowness. The evidence supporting the conclusions is solid, although validating their behavioural measures and considering the relationship between signatures from different brain regions would have strengthened the study. The work will be of broad interest to movement control, movement disorders, and brain stimulation fields.
-
Reviewer #1 (Public Review):
Authors aimed to decode signatures linked to tremor, slowness and effective motor control using different types of signals acquired from a group of Parkinson's disease patients during deep brain stimulation surgery. They were able to identify distinct frequency bands which corresponded to different symptoms and conclude that multi frequency band and cortical decoding surpass single frequency band and subthalamic nucleus-based decoding.
The main strength of the study is the recording types used to decode symptoms emerging during the same experimental task: authors leveraged micro and macro level recordings from the subthalamic nucleus and ECoG recordings from the motor cortex, enabling them to provide decoding performance across distinct recording scales and from two critical structures linked to Parkinson's …
Reviewer #1 (Public Review):
Authors aimed to decode signatures linked to tremor, slowness and effective motor control using different types of signals acquired from a group of Parkinson's disease patients during deep brain stimulation surgery. They were able to identify distinct frequency bands which corresponded to different symptoms and conclude that multi frequency band and cortical decoding surpass single frequency band and subthalamic nucleus-based decoding.
The main strength of the study is the recording types used to decode symptoms emerging during the same experimental task: authors leveraged micro and macro level recordings from the subthalamic nucleus and ECoG recordings from the motor cortex, enabling them to provide decoding performance across distinct recording scales and from two critical structures linked to Parkinson's pathophysiology. This allowed the authors to contrast rhythm-based signatures and timescales of Parkinson's disease motor symptoms.
The primary weakness is the level of description used to describe key methods which makes it difficult to unpack the results: authors should pay particular attention to validating and justifying metrics used for assessing behaviour (e.g., tremor, slowness, and effective motor control). Also, the relationship between behavioural measures and UPDRS scores should be further justified. For instance, (1) what is the definition of tremor amplitude probability density in the absence of tremor and what is its relationship to relevant subcategories of clinical tremor severity?; (2) why did the authors link tremor while performing a task to UPDRS rest tremor scores? ; (3) why did the authors opt for normalised cursor speed as a metric for slowness?; (4) Are there any implications of this normalisation when exploring slowness across participants? Authors consider cortical and subthalamic recordings separately: if these recordings were acquired simultaneously, analysing the relationship between the two signals (i.e., envelope, phase, phase-amplitude) would significantly improve the paper.
Authors aimed to decode signatures linked to different symptoms of Parkinson's disease. Results support their primary conclusions that cortical decoding performs better than subthalamic decoding and that using a multi-dimensional feature space improves the performance of the decoder. The paper and data generated will contribute to movement control, movement disorders, and brain stimulation fields.
-
Reviewer #2 (Public Review):
The present study aimed to demonstrate the utility of brain signal decoding for the differentiation of asynchronous motor signs in Parkinson's disease (PD). To this end, thirty-one PD patients undergoing deep brain stimulation electrode implantation were recruited to participate in an intraoperative motor task. Task performance was compared to extra-operative experiments in healthy subjects. Neural activity and movement traces were segmented into 7-second windows and attributed tremor and slowness measures. To integrate the two symptom domains an additional decoding state termed effective motor control was introduced, which represented the absence of symptoms. Support vector machine regression was used as the model of choice that was trained on individual recording sessions within subjects. All decoding …
Reviewer #2 (Public Review):
The present study aimed to demonstrate the utility of brain signal decoding for the differentiation of asynchronous motor signs in Parkinson's disease (PD). To this end, thirty-one PD patients undergoing deep brain stimulation electrode implantation were recruited to participate in an intraoperative motor task. Task performance was compared to extra-operative experiments in healthy subjects. Neural activity and movement traces were segmented into 7-second windows and attributed tremor and slowness measures. To integrate the two symptom domains an additional decoding state termed effective motor control was introduced, which represented the absence of symptoms. Support vector machine regression was used as the model of choice that was trained on individual recording sessions within subjects. All decoding targets from each neurophysiological modality reached significant prediction performances. This represents an important milestone in the current state of research towards machine learning-based intelligent adaptive deep brain stimulation.
Strengths
1. The present analysis is among the first to demonstrate the potential utility of brain signal decoding for the differentiation of asynchronous motor symptoms in Parkinson's disease. In the future, such approaches may be adopted in clinical brain-computer interfaces that can adapt stimulation in real time to concurrent therapeutic demand.
2. The effort from the research team and patients to acquire this important dataset is commendable. The time pressure in the operation room combined with the current trend of asleep surgery for deep brain stimulation makes such data very rare.
3. No relevant difference in decoding performance was found for subthalamic micro vs. macroelectrode recordings. This has practical significance because current sensing-enabled deep brain stimulation implants only allow for macro-recordings, which according to this study has no severe disadvantage over microelectrode recordings for movement decoding. Note that this question could only be answered in the intraoperative setting, which on the other hand can have disadvantages further described below.
4. Beyond the subthalamic nucleus, the authors corroborate the superiority of electrocorticography over subthalamic activity for movement and symptom decoding in Parkinson's disease. This provides further evidence that additional sensing electrodes may complement the subthalamic signals for adaptive deep brain stimulation.
5. Finally, the idea of decoding the presence of an effective motor state is creative and may inspire future developments in adaptive stimulation control algorithms.
Weaknesses
(Note that I take more words for weaknesses, not because they outweigh the strengths, but because I want to justify my criticism in more detail)
1. One inherent limitation of this study is the intraoperative setting, which demands the patients' skull be fixed to the stereotactic frame. This setting is not naturalistic per se and likely comes with additional perturbations in the brain states that are recorded. Thus, the generalization to real-world scenarios is limited. Given the unique opportunity to record invasive brain signals in humans, this limitation has to be accepted and should be taken into account for the interpretation of the results. As mentioned in the strengths, this is currently the only setting that allows for a comparison of micro- and macroelectrode recordings for brain signal decoding.
2. Similarly, the medication state is defined by the intraoperative scenario, as deep brain stimulation implantations are performed after the withdrawal of dopaminergic medication in the so-called dopaminergic OFF state. In this state, PD symptoms are aggravated, which is used clinically to provide a more reliable assessment of deep brain stimulation-induced symptom alleviation. This may also lead to an overestimation of decoding performances as the difference between the absence and presence of PD motor signs in the dopaminergic medication ON state during activities of daily living could be more nuanced.
3. The task design is very interesting as it allows for a continuous definition of symptom severity and motor performance. The comparison to healthy subjects demonstrated clearly higher tremor scores in PD but no significant differences in movement velocity (depicted as trending but p>0.2). This is somewhat unexpected as slowness of movement, also called bradykinesia, is a defining symptom of Parkinson's disease (PD). By definition, this symptom is present in all PD patients, also indicated in the clinical scores shown in the present study. Action tremor, i.e. the presence of tremulous muscle activity during motor performance, is comparatively rare. To support the clinical relevance of the movement tremor observed during the task, the authors show a correlation with the "resting tremor" score from the clinical assessment. It is unclear to me why resting, instead of action tremor scores are shown, as both are part of the clinical assessment (Unified Parkinson's disease rating scale - UPDRS part III). Ultimately, even though resting tremor is significantly more common in Parkinson's disease, not all patients of the current cohort had resting tremor (as indicated in the clinical score correlation). Thus, it remains somewhat puzzling how precise the 3-8 Hz activity actually captures tremor vs. motor noise or inaccuracy. A more fine-grained analysis comparing patients with clinically diagnosed action tremor (as defined by preoperative UPDRS assessment) and without tremor could have helped to support the clinical claims on symptom-specific decoding. On the other hand, the lack of a significant difference in the slowness of movement in the patient cohort relative to healthy controls questions the ability of the task to capture this symptom. Here, I am not sure whether the normalization procedure may have an influence on the comparability. Finally, movement velocity is an easy target that is distributed across a spectrum, so despite the lack of a significant difference in the healthy cohort, I am relatively confident that the decoding of movement slowness in the present cohort is clinically meaningful.
4. Overall, the pathophysiological framework is well placed in the current state of literature, while almost the entire field of brain signal decoding for adaptive deep brain stimulation was neglected. Successful decoding to address Parkinson's and essential tremor (another disorder with more common action tremor) was achieved by multiple groups in impactful studies representing more naturalistic extraoperative or fully embedded settings (Hirschmann et al., 2017, He et al., 2021, Opri et al., 2021). Additionally, other symptoms, like gait disturbances have been the target of machine learning analyses more recently (Louie et al., 2022 and Thenaisie et al., 2022). Here, the manuscript appears to avoid a discussion of the present endeavour in comparison to the current state of the field. One of our own studies has provided the first demonstration of the superiority of electrocorticography over subthalamic LFP for movement decoding, which I am happy to see replicated for the first time in the present manuscript. Importantly, the referenced study showed modality-dependent model performances, with gradient-boosted decision trees performing significantly better than linear models for electrocorticography, while Wiener filters have been repeatedly shown to perform well for subthalamic local field potentials (e.g. see Shah et al., 2018 IEEE Trans Neural Syst Rehabil Eng). The present study does not compare different machine learning architectures. Thus, decoding performances could potentially be further improved with more refined computational approaches. A more thorough overview of the literature from the many laboratories that are invested in this research across the globe would have improved the interpretation with respect to the broader impacts of the present manuscript.
5. The authors also present analyses of the spatial localization of relative decoding performances. They demonstrate higher tremor decoding performance in the dorsolateral subthalamic nucleus and higher decoding performance for the slowness of movement in the more central and ventral subthalamic regions. The authors interpret this as potential evidence to support clinical decision-making for optimized stimulation control of these symptoms at the respective locations. This is overly speculative and currently not backed by the data. First of all, the results only show the contrast of tremor vs. slowness of movement and not each individually. Thus, the spatial peak with each symptom domain could be very similar, e.g. in the dorsolateral STN, but a reversal of the difference only occurs at relatively low performances, e.g. in the ventral STN. Thus, showing both spatial distributions individually could be more informative. However, the claim that this could also be used to adjust stimulation location to alleviate the respective target symptoms is by no means backed by the data and remains an interesting speculation.
6. Finally, as in many brain signal decoding studies, the presented decoding performances are relatively low. The authors decided to present linear correlation metrics as Pearson's r values. These values are by definition higher than the commonly chosen Coefficient of determination or R² that provides a more interpretable performance metric. The amount of variance in the symptom scores that could be explained by the models ranged between 10% and 30% at a temporal resolution of 7 seconds. Moreover, the validity of the linear score is not entirely clear as Pearson's r can be heavily biased by non-normal distributions which were not assessed or at least not reported for the performance evaluation. These considerations do not severely limit the validity of the results themselves, as the authors have convincingly shown that significant decoding performances are possible and other studies in this field range in similar performance ranks. However, this point should remind us that a short-term clinical adoption of such methods is not yet in sight and further research is warranted. Before machine learning-based clinical computer interfaces can reach the clinical routine, the field has to work on more refined methods. In my opinion, the field will have to provide robust decoding performances with R² > 0.8 without patient-specific training to get into the realm of widespread clinical adoption.
-
Reviewer #3 (Public Review):
In this manuscript, the authors examine microelectrode and macroelectrode recordings from the human STN, as well as electrocorticography from the sensorimotor cortex in order to examine the neurophysiological biomarkers underlying tremor and bradykinesia. This is an important and timely topic, as the detection of such biomarkers can have implications for developing effective closed-loop DBS devices. Currently, there is some uncertainty as to which biomarkers may be relevant for which particular symptoms. Here the authors examine signals recorded from multiple depths within the STN and regress those signals onto behavioral measures of tremor and slowness as captured using a novel behavioral paradigm in which patients track movements on a screen in the intraoperative setting. This group has published on this …
Reviewer #3 (Public Review):
In this manuscript, the authors examine microelectrode and macroelectrode recordings from the human STN, as well as electrocorticography from the sensorimotor cortex in order to examine the neurophysiological biomarkers underlying tremor and bradykinesia. This is an important and timely topic, as the detection of such biomarkers can have implications for developing effective closed-loop DBS devices. Currently, there is some uncertainty as to which biomarkers may be relevant for which particular symptoms. Here the authors examine signals recorded from multiple depths within the STN and regress those signals onto behavioral measures of tremor and slowness as captured using a novel behavioral paradigm in which patients track movements on a screen in the intraoperative setting. This group has published on this paradigm previously, and here they now use support vector regressions to examine how the physiological data relates to these behavioral measures. In brief, they find that tremors and bradykinesia (slowness) correlate with different neural signatures from different locations. Overall, the results seem well supported, and the methods and statistical tests are sound.
-
