Decoding gripping force based on local field potentials recorded from subthalamic nucleus in humans

Abstract

The basal ganglia are known to be involved in the planning, execution and control of gripping force and movement vigour. Here we aim to define the nature of the basal ganglia control signal for force and to decode gripping force based on local field potential (LFP) activities recorded from the subthalamic nucleus (STN) in patients with deep brain stimulation (DBS) electrodes. We found that STN LFP activities in the gamma (55-90 Hz) and beta (13-30m Hz) bands were most informative about gripping force, and that a first order dynamic linear model with these STN LFP features as inputs can be used to decode the temporal profile of gripping force. Our results enhance the understanding of how the basal ganglia control gripping force, and also suggest that deep brain LFPs could potentially be used to decode movement parameters related to force and movement vigour for the development of advanced human-machine interfaces.

Keywords: basal ganglia; beta oscillation; gamma oscillation; gripping force; human; local field potentials; neuroscience.

Figure 1.. Force-effort scaling and spectra of average power changes relative to pre-movement baseline for two groups of electrodes.

(A) Trajectory of measured force from one exemplar subject. (B) From one group of electrodes (n = 9), a significant reduction of power in the beta band (13–30 Hz) and increase in power in the broad gamma band (55–90 Hz) was observed during gripping. (C) In another group of electrodes (n = 9), significant simultaneous modulation was absent in the beta and gamma band with movement, and there was instead an increased power across the low frequency band during gripping. Trajectories of force (D) and force yank (G) for Group one show that the stable force during the holding phase (1–2 s after cue) as well as the peak force yank in the force initialisation phase scaled well with self-rated effort (SRE). In group 2, the stable force (E) and force yank (H) did not scale with effort as well as Group 1. Group two had significantly lower correlation coefficients between stable force and SRE (F) and between the peak force yank and SRE (I) compared with group 1 (p<0.0001), indicating some impairment in the scaling of force with effort. Time 0 indicates the onset of the cue to start a grip in B–E, G and H. Note that data from three electrodes are excluded. Of these, two had significant modulation in the beta band but not in the gamma band, and one had significant modulation in the gamma band, but not in the beta band. DOI: http://dx.doi.org/10.7554/eLife.19089.003

Figure 2.. Force prediction performances of different models evaluated in terms of within-trial correlation (A), RMSE (B) and BIC (C).

The filled dots and shaded bars show the median and range across all STNs; the open circles and stars show the data for each individual STN. The red dots and bars show performance in predicting high effort forces, while using data from low effort trials for model fitting; the blue dots and bars show performance in predicting low effort forces, while using data from high effort trials for model fitting. (C) The total BIC values combing the force predictions for low effort and high effort for all tested models. The filled dots and shaded bars show the median and range across all STNs; the stars show the data for each individual STN (some overlap). Models 1–3 use beta and gamma ERS as model inputs; Models 4–5 use activities from all three frequency bands (alpha, beta and gamma) as model inputs; Models 6–8 use activities from a single frequency band (alpha, beta and gamma, respectively) as model inputs. DOI: http://dx.doi.org/10.7554/eLife.19089.007

Figure 3.. Fitting and predicting performance of the model for predicting force averaged across multiple trials.

(A) The fitted model based on data from low effort trials for one exemplar STN and the contralateral hand. (B) The fitted model was used to predict the average force for high effort trials for the same STN and contralateral hand. (C) The fitted (dashed lines) and predicted force were compared against the measured force for the other 8 STNs in which consistent movement-related modulations in both beta and gamma bands were observed. Predicted force traces for high effort trials were derived from the model fitted to data from low effort trials and vice versa. Time 0 indicates the onset of the cue to start a grip in all plots. DOI: http://dx.doi.org/10.7554/eLife.19089.009

Figure 4.. Predicting force profile of individual grips based on beta and gamma activities from STN LFP (one exemplar subject).

(A) Time-evolving power spectrum of the bipolar STN LFP channel used for decoding force. (B) The predicted force (in red) compared with the measured force (in black). ** indicates the trial where STN LFP predicted increased force but with no measured force from the dynamometer. Grips are concatenated in A and B. (C) Distribution of the within-trial correlation coefficient (WithinR) between predicted force and measured force, with the dashed blue line the median value of the WithinR for all trials. (D) Scatter plot between the predicted stable force (average force during the second of holding phase) and measured stable force for all tested trials. The correlation coefficient between the predicted and measured stable force across trials was 0.815 for this subject. The regression slope of 0.96, which is close to 1, shows that there is no systematic under-estimation. The black lines show the regression line and 95% confidence interval. DOI: http://dx.doi.org/10.7554/eLife.19089.011

Figure 5.. STN LFP features predict gripping force profile generated by the contralateral hand.

(A) The correlation coefficients between the measured stable force and the predicted stable force were higher for the force generated by the contralateral hand than that by the ipsilateral hand. There was no significant difference when different models based on both beta and gamma activities from STN LFP were used. The dots and bars show the median value and the range of values for different STNs. ** indicate a significant difference in the prediction performance when the LFPs from the ipsilateral STN was used for decoding. (B) The histogram of the within-trial correlation coefficients between predicted force and measured force (WithinR) for the contralateral hand considering all the trials and all the STNs. (C) Cumulative distribution function (CDF) of the WithinR for the force generated by the contralateral hand (solid lines) and the ipsilateral hand (dashed lines). The CDF indicates the probability that WinthinR has a value less than or equal to a certain value on the x-axis. Data presented in this figure are for all the STNs in which significant modulations were observed in both the beta band and gamma band. DOI: http://dx.doi.org/10.7554/eLife.19089.013

Figure 6.. Factors affecting the gripping force prediction performance.

The median values of the WithinR (A) and stable force correlation coefficients (B) increased with the average movement- related modulation in the gamma band. Each dot is the data for one STN and the blue line shows the exponential fit of the data ($y=a\bullet e^{-bx}+k$, p<0.001 for the fitting). The median values of the WithinR (C) and stable force correlation coefficients (D) show a trend of increasing with average movement related desynchronization in the beta band. The blue lines show a linear fitting, but the fits were not significant. In this Figure, we consider all the STNr with significant movement-related modulations, whether either in one or other, or both frequency bands of interest (N = 12). DOI: http://dx.doi.org/10.7554/eLife.19089.015

Figure 7.. Validation of the models for force prediction on an independent patient group during maximal effort gripping.

(A) Average power change in the STN LFP activity associated with the gripping movement. The power change is relative to the average over a 1 s period pre-cue. Time 0 is the timing of cue onset. (B) Median WithinR for individual STN and contralateral hands, * indicates data for each individual STN and contralateral hand. (C) Histogram and (E) cumulative distribution function (CDF) of the WithinR for all the 397 individual trials across all the 20 STNs. (D) BIC analysis showed that Model 4 considering alpha, beta and gamma power changes significantly improved force prediction compared to Model 2 during maximal effort gripping. * indicates p<0.05 using a paired t-test. DOI: http://dx.doi.org/10.7554/eLife.19089.016

Figure 8.. Implications of reduced movement related modulation in beta and gamma band activity in STN LFP with reduced dopaminergic input.

(A) The range of forces that can be generated will be reduced if the scale between the STN encoding signal and the force is to remain the same. This will lead to unscaled, bradykinetic force generation. (B) The scale between the STN encoding signal and the force will be increased if the range of force that can be generated is to be kept similar. This will lead to abnormally high force generation and more coarse force control. DOI: http://dx.doi.org/10.7554/eLife.19089.018