Identifying controllable cortical neural markers with machine learning for adaptive deep brain stimulation in Parkinson's disease

Abstract

The identification of oscillatory neural markers of Parkinson's disease (PD) can contribute not only to the understanding of functional mechanisms of the disorder, but may also serve in adaptive deep brain stimulation (DBS) systems. These systems seek online adaptation of stimulation parameters in closed-loop as a function of neural markers, aiming at improving treatment's efficacy and reducing side effects. Typically, the identification of PD neural markers is based on group-level studies. Due to the heterogeneity of symptoms across patients, however, such group-level neural markers, like the beta band power of the subthalamic nucleus, are not present in every patient or not informative about every patient's motor state. Instead, individual neural markers may be preferable for providing a personalized solution for the adaptation of stimulation parameters. Fortunately, data-driven bottom-up approaches based on machine learning may be utilized. These approaches have been developed and applied successfully in the field of brain-computer interfaces with the goal of providing individuals with means of communication and control. In our contribution, we present results obtained with a novel supervised data-driven identification of neural markers of hand motor performance based on a supervised machine learning model. Data of 16 experimental sessions obtained from seven PD patients undergoing DBS therapy show that the supervised patient-specific neural markers provide improved decoding accuracy of hand motor performance, compared to group-level neural markers reported in the literature. We observed that the individual markers are sensitive to DBS therapy and thus, may represent controllable variables in an adaptive DBS system.

Keywords: Adaptive deep brain stimulation; Brain-computer interface; Deep brain stimulation; Machine learning; Neural marker.

Fig. 1

Left: Illustrative EEG layout used in sessions $S_2^{+2,+3}$. The EEG electrodes avoided incision sites used for DBS electrode implantation. In sessions $S^c$, a full EEG layout was used. Right: The experimental setup and design of the copy-draw test, as composed by (a) single trial trace consisting of three (b) trace atoms, (c) get-ready box, and (d) starting point of the trace.

Fig. 2

Processing pipeline of the EEG data. The lower row represent that main pipeline, the upper row is performed for identifying biological artifactual epochs and ICA components (by visual inspection), which are removed from the data in the main pipeline.

Fig. 3

Hand motor performance z extracted in task T1 and the area under the ROC curve (AUC) describing the separability between DBS-on/-off. For all sessions, AUC scores are significantly above chance level, tested using a bootstrapping procedure with 200 label shuffles at a significance level of $p<0.05$ (uncorrected).

Fig. 4

Per patient and session, accuracies (y-axes) achieved for the decoding of motor performance (green), and of DBS-condition (pink) are provided relative to varying central frequencies $f_c$ (x-axes). Accuracies for different hyperparameter configurations of $f_c$ and $f_w$ have been grouped into 2 Hz bins and averaged. Circle size represents the percentage of evaluations for a given 2 Hz bin which achieve an accuracy above chance level. Chance level is indicated by solid lines.

Fig. 5

Distribution of decoding performance in the regression task (task T2), according to the stimulation condition of the trials in the test folds. In some specific cases, there is an apparent difference in performance according to the homogeneity of the stimulation condition in the test set, however, a global trend cannot be identified.

Fig. 6

DBS-induced changes in motor performance (AUC during extraction of motor performance) vs. the highest accuracy achieved in the decoding of motor performance from EEG signals.

Fig. 7

Informative components of neural origin obtained for the decoding of motor performance and of DBS condition. Each panel contains: frequency spectrum of the spatially filtered data (top-left); time-locked dynamics of the spatially and frequency-filtered data wrt. to the beginning of the trial (bottom-left); spatial pattern of the component (right). Vertical black lines in the bottom-left plot mark the beginning and end of the copy-draw trial. Unless indicated, components shown were found informative for both motor performance and DBS condition. For $S_2^c$, the spatial patterns of components indicated biological artifactual origins and thus, are not shown. Note that the polarity of the spatial patterns shown is arbitrary.

Fig. 8

Continued: informative components of neural origin obtained for the decoding of motor performance and DBS condition.

Fig. 9

Comparison between motor performance decoding using spatio-spectral decomposition of the fixed theta-, alpha-, and beta-bands (x-axis) vs. the overall best performing data-driven NMs identified with our approach (y-axis), not limited to any frequency band. Sessions above the dashed line indicate better performance of the data-driven NMs. Statistical significance is indicated by filled markers and was determined with Bonferroni corrected Wilcoxon signed-rank test at p-value of 0.05.

Fig. 10

In an adaptive DBS system, motor performance is estimated by a decoder for neural markers (red) from brain signals. Its output informs a controller, which decides on the ongoing stimulation intensity. Using the neural marker information, the controller does not require any active polling of motor performance—this is done during calibration only (in yellow).

Fig. 11

Left: Average power spectral density of stimulation artifacts computed from LFPs of the STN for a representative session ($S_1^{+2}$). Right: Corresponding average of the stimulation artifacts induced on the EEG channels.

Fig. 12

Left: CSP spatial patterns computed on the $[128-132]$ Hz band, representing the topological features of the underlying stimulation artifacts for sessions $S^c$ . The topological features displayed are also found in the harmonics of the stimulation frequency (260 Hz). Right: Spectral density of the signals corresponding to the aforementioned spatial filters, in the frequency band considered for tasks T2 and T3. Note that the polarity of the spatial patterns shown is arbitrary.