Electrocorticography is superior to subthalamic local field potentials for movement decoding in Parkinson’s disease

Curation statements for this article:
  • Curated by eLife

    eLife logo

    Evaluation Summary:

    This manuscript will be of interest to the motor neuroscience, movement disorder, human electrophysiology, and brain computer interface communities. It examines the effectiveness of signals recorded from the subthalamic nucleus (STN) and along sensorimotor regions of the cortex for decoding simple movements in patients with Parkinson's disease. Additionally, a relationship between symptom severity and decoding performance is identified. With the recent advent of implantable closed-loop stimulators for neurological conditions such as Parkinson's disease, this paper addresses current knowledge gaps that may inform both surgical and engineering considerations for optimizing these new types of therapies.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 agreed to share their name with the authors.)

This article has been Reviewed by the following groups

Read the full article

Abstract

Brain signal decoding promises significant advances in the development of clinical brain computer interfaces (BCI). In Parkinson’s disease (PD), first bidirectional BCI implants for adaptive deep brain stimulation (DBS) are now available. Brain signal decoding can extend the clinical utility of adaptive DBS but the impact of neural source, computational methods and PD pathophysiology on decoding performance are unknown. This represents an unmet need for the development of future neurotechnology. To address this, we developed an invasive brain-signal decoding approach based on intraoperative sensorimotor electrocorticography (ECoG) and subthalamic LFP to predict grip-force, a representative movement decoding application, in 11 PD patients undergoing DBS. We demonstrate that ECoG is superior to subthalamic LFP for accurate grip-force decoding. Gradient boosted decision trees (XGBOOST) outperformed other model architectures. ECoG based decoding performance negatively correlated with motor impairment, which could be attributed to subthalamic beta bursts in the motor preparation and movement period. This highlights the impact of PD pathophysiology on the neural capacity to encode movement vigor. Finally, we developed a connectomic analysis that could predict grip-force decoding performance of individual ECoG channels across patients by using their connectomic fingerprints. Our study provides a neurophysiological and computational framework for invasive brain signal decoding to aid the development of an individualized precision-medicine approach to intelligent adaptive DBS.

Article activity feed

  1. Evaluation Summary:

    This manuscript will be of interest to the motor neuroscience, movement disorder, human electrophysiology, and brain computer interface communities. It examines the effectiveness of signals recorded from the subthalamic nucleus (STN) and along sensorimotor regions of the cortex for decoding simple movements in patients with Parkinson's disease. Additionally, a relationship between symptom severity and decoding performance is identified. With the recent advent of implantable closed-loop stimulators for neurological conditions such as Parkinson's disease, this paper addresses current knowledge gaps that may inform both surgical and engineering considerations for optimizing these new types of therapies.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 agreed to share their name with the authors.)

  2. Reviewer #1 (Public Review):

    This paper evaluates the decoding performance of sensorimotor electrocorticography (ECoG) and subthalamic nucleus (STN) local field potentials (LFPs) during a grip-force task in patients with Parkinson's disease. ECoG signals showed better decoding performance compared to STN LFPs or the combination of the two. Gamma band power seemed to provide the most information across power bands. Individuals with less impairment showed the best decoding performances, which was also impacted by both structural and functional connectivity with the chosen electrode contact. Overall, the authors took a comprehensive approach for evaluating decoding performance using a variety of algorithms in a pseudo real-time fashion to establish potential superiority of ECoG over the sensorimotor cortex compared to STN LFPs for using power to predict grip force production.

    Strengths:

    Despite being a retrospective analysis, the authors used appropriate methodological considerations to evaluate pseudo real-time decoding performance with algorithms and methods that could be translated to a true real-time performance. The authors evaluated power bands relevant to the technological capabilities of current adaptive neurostimulators. The authors provided sufficient support for both the superiority of ECoG over STN LFPs, as well as signals from the contralateral side compared to ipsilateral signals. These findings held true for both different timings (i.e., the duration of time over which to calculate a given band's power relative to movement onset) and algorithm complexity.

    The authors also investigated the contribution of factors other than just the neural signals themselves on decoding performance. The finding that impairment level was related to decoding performance will be relevant for any potential clinical utility of these types of decoding algorithms. Similarly, it is well-known that electrode location will impact the relevance of a given electrode's signal. The author's attempt to address this in a multi-faceted fashion by first evaluating the impact of electrode distance to relevant anatomical markers (hand-knob region of motor cortex and dorsolateral STN), and then also evaluating the roles of both structural and functional connectivity profiles for a given contact location. These significant findings may help explain the wide spread observed in decoding performance.

    Although the focus of this paper is in the context of PD, the proposed methods are applicable to many domains of brain computer interfaces and neural signal decoding.

    Weaknesses:

    A wide-range of decoding performance is seen across participants with a group of good performers and a group of low-performers. The findings from the secondary analyses on impairment levels and electrode location/connectivity may explain some of these differences, but it is unclear to what extent or if other factors are at play.

    The cohort is notably small, especially considering the heterogeneity of electrode location. Although this does not limit the major finding of the superiority of sensorimotor ECoG over STN signals, it does limit the ability to understand the observed individual differences in decoding performance.

    The authors are unable to fully rule-out that the superiority of ECoG signals is not dependent on the task performed (i.e., grip force). The authors claim that the findings support the utility of additional ECoG in adaptive DBS research for PD patients. Although it is certainly expected that sensorimotor ECoG would provide a richer signal compared to STN LFPs due to both signal size, complexity, and relation to movement, the data in this paper is inherently limited to the grip task performed.

    Conclusion:

    The methods used to evaluate sensorimotor ECoG and STN LFP signals for grip force production are of interest for both the PD field (e.g., aDBS and neurophysiological underpinnings of behavior and impairment) as well as the generalized field of brain-computer interfaces. The authors address how a multitude of factors (timing of signal, location of recording electrodes, frequency bands used, impairment level etc.) may impact decoding performance, which can inform/guide future work. The current work shows strong evidence for the superiority of sensorimotor ECoG compared to STN LFPs for decoding grip force production in PD.

  3. Reviewer #2 (Public Review):

    Merk et al., compare grip force decoding performance between cortical ECoG electrodes and subthalamic LFP and find that electrodes over cortical regions perform better. They first compare a simple linear regression model, and then use several more sophisticated techniques to decode grip performance for each electrode for both contra and ipsi movements. Overall the claim that ECoG electrodes decode grip force with higher accuracy than subthalamic LFP seems supported with their data, although there are some inherent limitations of the clinical data that need to be addressed if not with data than in the discussion section. In addition, they find that decoding performance is negatively correlated with PD impairment and they use connectivity models to identify if decoding performance is related to connectivity profiles.

    I appreciate that this paper uses several different decoding techniques and attempts to decode grip force for contra and ipsi movements for each electrode. The main result of this paper is that neural signals from ECoG electrodes are superior to subthalamic LFP for movement decoding. Based on the analyses that the authors provide, these results seem to be of potential interest for clinical researchers interested in adaptive deep brain stimulation (aDBS) and basic science researchers interested in motor control. Although the difference in grip force decoding appears quite large, there are a couple limitations that I think the authors could address to make the paper even stronger.

    The first limitation when comparing cortical and subthalamic electrodes is that the size and structure of the probe may be different. This means that instead of comparing apples to apples, it is more like comparing apples to oranges. This does not completely undermine the result because the difference in decoding between the areas, even given experimental differences, is likely to be of interest to clinical researchers studying DBS. If the surface area of the electrode is different between the two regions, then this could be a factor in decoding performance that does not have to do with brain region. Additionally, the electrodes in the subthalamus nucleus are circular, which are likely targeting very different neural populations across the probe within the small nucleus, which is different from the cortical electrodes which are on the surface targeting neural populations which are adjacent. Both of these factors (e.g. size and shape) could contribute to differences in decoding performance regardless of brain region. I did not see details of the electrodes in the method section, but this would be important to report as surface area is related to the number of neurons/dendrites summing to create the LFP, and this might lead to qualitatively different results for something like hand gripping irrespective of area. Similarly, with the shape of the electrode. These details will be an important addition to the paper and something that others can continue to investigate (e.g., researchers who have different size or shape of electrodes in the STN). I am sympathetic that this is not a variable that the researchers can change given the clinical nature of DBS, but the surface area of electrodes in each area should be mentioned in the method section, and if the surface area of the electrodes are different, then it should also be mentioned as a limitation in the discussion section. Nonetheless, the results are likely to be of interest for clinical researchers, but they would need these details in order to compare to their own DBS system (there are now directional leads which have more electrodes and thus smaller surface area).

    The second possible limitation is whether you have fully explored the neural feature space. Although the cutoffs for frequency bands remain somewhat arbitrary, your selection of frequency bands seems very reasonable and seems to cover all the possibilities. One suggestion I have is that you also include the time domain data as a feature along with your frequency bands. Some papers have shown pretty good decoding with this feature - sometimes called the local motor potential. Here are some papers which discuss this feature in more detail. This could be an interesting addition especially if it performs well as it requires little preprocessing for studies doing online preprocessing and decoding.

    Flint, R. D., Wang, P. T., Wright, Z. A., King, C. E., Krucoff, M. O., Schuele, S. U., ... & Slutzky, M. W. (2014). Extracting kinetic information from human motor cortical signals. Neuroimage, 101, 695-703.

    Mehring, C., Nawrot, M. P., de Oliveira, S. C., Vaadia, E., Schulze-Bonhage, A., Aertsen, A., & Ball, T. (2004). Comparing information about arm movement direction in single channels of local and epicortical field potentials from monkey and human motor cortex. Journal of Physiology-Paris, 98(4-6), 498-506.

    Schalk, G., Kubanek, J., Miller, K. J., Anderson, N. R., Leuthardt, E. C., Ojemann, J. G., ... & Wolpaw, J. R. (2007). Decoding two-dimensional movement trajectories using electrocorticographic signals in humans. Journal of neural engineering, 4(3), 264.

    Given how similar the descriptive power plots are, I am surprised that low gamma has much larger weights compared to high gamma or HFA. It looks like you aren't using regularization for your linear regression model. If your features (band pass filters) are highly correlated, the interpretation of the weights might not be meaningful. Have you thought about using ridge regression or lasso to deal with your seemingly highly correlated features? If not then I don't believe it makes sense to try and interpret the weights. It looks like you do use regularized regression later, but looking at the method section for your linear regression model there is no regularization term - so based on that it seems like for this first section it is just standard linear regression. I would suggest also using regularized regression for these analyses as interpreting the weights of linear regression with highly correlated features may be problematic.

    The correlation with decoding performance and motor PD impairment is intriguing and I think this analysis and the result is of value to both clinical and basic researchers.

    Although not dependent on your main claim, I had a difficult time understanding the logic and the methods of your last section which relates decoding performance with connectivity maps. For example, after reading the methods section, I was still unclear how you determined if a fiber was significant or not. I believe that this section needs more detail and clarity before publication. For example, you have analyses for structural and functional connectivity, but for the functional connectivity I could not find anything in the method section about what the patient was doing when this was computed - were the patients at rest, were they doing the same gripping task? These details are important for understanding the analyses and interpretation.

  4. Reviewer #3 (Public Review):

    This work evaluates the effectiveness of signals recorded in the subthalamic nucleus (STN) and along sensorimotor regions of the cortex for decoding simple movements in patients with Parkinson's disease. The authors present this motor decode as a potential control signal for adaptive deep brain stimulation. A structured machine learning approach is presented that investigates the value of different sensor recording locations, signal components (frequency bands), and decoding model architectures. Additionally, a relationship between symptom severity and decoding performance is identified. With the recent advent of implantable closed-loop stimulators for neurological conditions such as Parkinson's disease, this paper addresses current knowledge gaps that may inform both surgical and engineering considerations for optimizing these new types of therapies.

    Strengths:

    The authors present a clean and principled model testing procedure with appropriate training, validation, and testing. In principle, this produces unbiased performance estimates for each model with its best possible configuration of hyperparameters. This provides a nice format in which hypotheses about different modelling aspects may be compared. The general pipeline may be built upon in future studies investigating control signals for closed-loop neuromodulation in a variety of neurological conditions. Data-driven machine learning approaches like those presented here allow for highly individualized settings of such neuromodulatory therapies.
    Although all analyses were performed offline, the data were processed in a manner that would be appropriate for the proposed use of providing control signals for real-time adjustment of deep brain stimulation.

    The analysis of feature importance (frequency band weights) in the linear models provides helpful intuition and sanity checks from a basic neuroscience perspective. By testing multiple decoder models, both linear and non-linear with varying levels of complexity, some intuition is also gained about the nature of the information content in these signals and how that differs across the recorded brain regions. The main result - that cortical recordings outperform STN recordings for decoding grip-force - is demonstrated in the more complex XGBOOST models.
    Consideration is given to multiple important factors contributing to decoder performance. A thoughtful combination of topographic features are considered, including the location of sensors with regards to specific neuroanatomical landmarks as well as connectivity profiles of the surrounding tissue. Investigating the relationship between connectivity profiles of the sensor locations - both structural and functional connectivity - is particularly intriguing. In principle, this could be incorporated into planning sensor implant locations.

    Weaknesses:

    A primary weakness of the paper is that some of the results and ideas are underdeveloped. While identifying a relationship between motor symptom severity and decoder performance is interesting on its own, there is no further investigation of this result for providing better intuition about why the relationship exists and what can or should be done about it. The authors discuss the potential impact beta bursts might have on decoder performance but do not pursue any analysis of their own data to determine whether those proposed hypotheses are supported.

    Similarly, it is reported that models combining signals from multiple sensors do not benefit from the added features, but there is very little follow-up or elaboration on this. There are many regularization and ensemble learning methods that ought to be tested before coming to strong conclusions based on this observation. It seems likely that given an optimal integration of these features, decoders would experience a performance benefit. While the specific XGBOOST models that they tested may have failed to benefit from larger feature combinations, the usefulness and generalizability of that knowledge is limited without more exhaustive testing and discussion.

    A Go/No-Go task is also used in the paper. This is potentially problematic, given that there may be modulation in the neural recordings associated with response inhibition. While the decoders are trained only to predict actual force production, it is possible that inhibitory responses could influence the decoder readout and impact the performance metrics on which every analysis of the paper hinges. It is also plausible that such inhibitory responses would asymmetrically affect the subthalamic decoders and the cortical decoders. Comparison of those two signal locations is the main topic of the paper, so this is a critical aspect to consider.