Invasive and Non-Invasive Neural Decoding of Motor Performance in Parkinson's Disease for Personalized Deep Brain Stimulation

Abstract

Decoding motor performance from brain signals offers promising avenues for adaptive deep brain stimulation (aDBS) for Parkinson's disease (PD). In a two-center cohort of 19 PD patients executing a drawing task, we decoded motor performance from electroencephalography (n=15) and, critically for clinical translation, electrocorticography (n=4). Within each session, patients performed the task under DBS on and DBS off. A total of 35 sessions were recorded. Instead of relying on single frequency bands, we derived patient-specific biomarkers using a filterbank-based machine-learning approach. DBS modulated kinematics significantly in 23 sessions. Significant neural decoding of kinematics was possible in 28 of the 35 sessions (average Pearson's $\text{r}= 0.37$). Our results further demonstrate modulation of speed-accuracy trade-offs, with increased drawing speed but reduced accuracy under DBS. Joint evaluation of behavioral and neural decoding outcomes revealed six prototypical scenarios, for which we provide guidance for future aDBS strategies.

Figures (17)

• Figure 1: Decoding DBS condition from behavior - CopyDraw ROC AUC. A: Behavioral decoding performance per subject and session (x-axis, session in subscript, numeric values correspond to days after surgery, subscript "c" stands for chronic sessions), sorted according to the CopyDraw ROC AUC (mean across LDA predictions in chronological cross-validation). Markers are colored by significance according to N=1000 bootstrap permutations. 95 % chance percentiles are shown with hollow markers. Session codes surrounded by a frame indicate sessions with invasive ECoG data. B: Signed SHAP values (feature importance) for LDA models trained on all data of the according sessions, sorted according to A. Positive SHAP values indicate larger features under DBS ON. C - E: Visualization of selected features and sessions to help with the interpretation of SHAP values. Features are presented as they are used within the LDA (clipped to three standard deviations per DBS condition and z-scored). All differences are significant (Mann-Whitney U test, $P<0.01$, two-sided). Abbreviations: ROC AUC, area under the receiver operating characteristic curve; SHAP Shapley, additive explanation; LDA, linear discriminant analysis; SEM, standard error of mean; STD, standard deviation.
• Figure 2: Task performance evaluation. A: Example template and trace visualizing components relevant for the task performance calculation. The target template (what should be drawn) is shown in blue for the parts that where matched by dynamic time warping (DTW) to the participants actual drawing (orange), and unmatched parts are shown in gray. Small gray lines visualize how trace samples are matched to template samples. B: Mean fraction of template covered (x-axis, the ratio of blue points to total template points in A), and mean average distance between trace and template. The means are calculated for each DBS condition and across all sessions which increased or decreased in task performance due to DBS separately. Gray lines indicate the $\pm$ one standard error of mean (SEM) range. C: Scatter plot comparing the DBS effect size on task performance (x-axis) with the CopyDraw ROC AUC (y-axis). Sessions with a significant effect size (Mann-Whitney U test, two sided) are shown with a black outline and text labels. Ordinary least-squares (OLS) fit lines are added for all sessions with positive effect size (increase under DBS) and all sessions with negative effect size (decrease under DBS) separately. Both sides show a significant correlation with r$=0.55$ ($P=0.014$) with $\text{R}^2=0.30$ for positive effect size and r$=-0.5$ ($P=0.049$) with $\text{R}^2=0.25$ for negative effect size. D: Session counts separated in positive (increase) and negative (decrease) effect size, separately for acute and chronic sessions. Sessions with no significant effect are shown with reduced opacity. E: Scatter plot comparing behavioral decoding of the DBS condition from CopyDraw score (x-axis) and from the same pipeline receiving only the task performance as a feature (y-axis). The dashed line shows x=y. Text labels are provided for session in which the task performance feature is more informative than the CopyDraw behavioral features for decoding of the DBS condition. Abbreviations: ROC AUC, area under the receiver operating characteristic curve;
• Figure 3: Neural decoding. A: Neural decoding of the CopyDraw score (diamond markers, y-axis) per subject and session (x-axis, session in subscript, numeric values correspond to days after surgery, subscript "c" stands for chronic sessions). Markers are colored by significance according to N=1000 bootstrap permutations. 95 % chance percentiles are shown with hollow markers. Sessions are sorted according to CopyDraw ROC AUC (see Figure \ref{['fig:behav_overview']}). Tick labels with a frame correspond to sessions with invasive ECoG data. B: Counts of selected relevant features per frequency band. Features were selected using the Minimum Redundancy Maximum Relevance (MRMR) feature selection algorithm. Sessions are sorted according to A. C: Performance of using the features (frequency and spatial filters) found in the regression pipeline (used in A) for the classification of the DBS condition (regression feature ROC AUC). Sessions are sorted as in A, with markers colored by significance according to N=1000 bootstrap permutations. 95 % chance percentiles are shown with hollow markers. D: Box plots for Pearson's correlation using the regression pipeline to predict the CopyDraw score vs. using the regression pipeline to predict the task performance. Both pipelines are evaluated using the chrono-CV with the according target variable (CopyDraw score or task performance). Decoding the CopyDraw score achieves significantly higher average correlation ($P < 0.01$). E: Scatter plot for neural correlation with the CopyDraw score (x-axis) and with the task performance (y-axis). The dashed line shows x=y. Text labels are provided for session in which the task performance can be decoded with higher correlation than the CopyDraw score. An ordinary least squares (OLS) fit shows significant correlation with r$=0.59$ ($P<0.001$), $\text{R}^2=0.35$, see OLS fit with 5 % confidence intervals (CI). Abbreviations: FBSPoC, filterbank source power comodulation; ROC AUC, area under the receiver operating characteristic curve; CV, cross-validation.
• Figure 4: Selected features showcasing features used for predicting the CopyDraw score. The top row show spatial patterns (in arbitrary units - AU) for three sessions with EEG and the location of the four contact ECoG strip for $\text{S17}_3$ over the Freesurfer average (fsaverage) brain, with the precentral gyrus highlighted. For the acute session ($\text{S12}_2$), the interpolation of the heatmap is limited to not show surface are which had no EEG electrodes applied due to the wound surgical wound area. The second row shows broad band power spectral density (PSD) for the spatially filtered signal (filters associated with the patterns shown above). For the ECoG feature, data corresponds to channel 1. Data is show as mean across trials for DBS ON and DBS OFF, with colored background for the standard error of mean (SME) and dashed outlines for +/- one standard deviation (STD). Gray backgrounds highlight the frequency range, the spatial filters were fitted from. Cluster permutation tests were applied with significance at $P<0.01$ shown by black horizontal lines. The third row showcases scaled band power of the spatially and frequency filtered signals as bars for each trial, colored according to DBS ON and OFF. These values are the input to the final linear regression layer in the decoding pipelines.
• Figure 5: Correlation between behavioral and neural decoding and examples of different outcome types. A - Correlation scatter between CopyDraw ROC AUC (y-axis) and mean correlation from neural decoding (x-axis). Ordinary least-squares (OLS) fit shows a significant correlation with $\text{r}=0.63$ ($P<0.001$) with $\text{R}^2=0.4$ and is drawn as dark gray line with 5 % confidence intervals (CI, gray area). Scatter points are colored according to the intracluster correlation coefficient (ICC). Sessions marked with a text label are presented as stereotypical examples in B-G. B-G visualizes the neural decoding correlations as used in A. Each scatter plot shows the actual CopyDraw scores (y-axis) vs. the CopyDraw score predictions from neural data (x-axis). Each point refers to a single trial and is colored according to the DBS condition---red for DBS ON and blue for DBS OFF. Like in A, the fitted OLS regression models are drawn as black lines. Marginal distributions are plotted as normalized histograms to the corresponding axes, colored according to the DBS condition. B shows all trials of $\text{S}10_c$---strong DBS effect on the CopyDraw score, which is also captured in the neural predictions. Both marginals show a clear bimodal distribution indicating a high ICC. C shows all trials of $\text{S}1_c$---an example of a continuous modulation of the CopyDraw score by DBS with good neural decoding performance. D shows all trials of $\text{S}5_2$---an example for a DBS modulation of the CopyDraw score, which still contains a lot of variability beyond the DBS modulation. No significant neural decoding was possible for this session. E shows all trials of $\text{S}12_2$---an example for a session in which no significant modulation of the CopyDraw score was possible, but in which the CopyDraw score could be predicted from neural signals. F shows all trials of $\text{S}3_c$---an example of a session in which a strong DBS effect on the CopyDraw score was observed, but which could not be decoded from neural signals. G shows all trials of $\text{S}15_2$---no effect of DBS on the CopyDraw score was observed and no neural decoding was possible.