A Framework for Adapting Deep Brain Stimulation Using Parkinsonian State Estimates

Abstract

The mechanisms underlying the beneficial effects of deep brain stimulation (DBS) for Parkinson's disease (PD) remain poorly understood and are still under debate. This has hindered the development of adaptive DBS (aDBS). For further progress in aDBS, more insight into the dynamics of PD is needed, which can be obtained using machine learning models. This study presents an approach that uses generative and discriminative machine learning models to more accurately estimate the symptom severity of patients and adjust therapy accordingly. A support vector machine is used as the representative algorithm for discriminative machine learning models, and the Gaussian mixture model is used for the generative models. Therapy is effected using the state estimates obtained from the machine learning models together with a fuzzy controller in a critic-actor control approach. Both machine learning model configurations achieve PD suppression to desired state in 7 out of 9 cases; most of which settle in under 2 s.

Keywords: Gaussian mixture models; Parkinson's disease; biomedical signal processing; deep brain stimulation (DBS); feature extraction; fuzzy control; state estimator; support vector machine.

Figure 1

A typical scheme for adapting DBS using PD state estimates.

Figure 2

Contour plot for state estimates over a feature space for the machine learning models. (A) Example feature space showing PD and non-PD examples for dataset C. (B) Probability density function (PDF) for PD and non-PD training examples in (A). (C) Contour plot for state estimates using SVM, with a range from 0 to 1 representing levels of severity from non-PD to PD for (A). (D) Contour plot for state estimates using GMM, with a range from 0 to 1 representing levels of severity from non-PD to PD for (A). The two features are, Feature 1 (21–26 Hz band) and Feature 2 (18–23 Hz band).

Figure 3

Modulating network used to simulate the effect of DBS on neuronal signals. (A) Basal-ganglia network model. (B) Frequency response for configuration with non-PD having higher amplitude in both bands. (C) Frequency response for a configuration with PD having higher amplitude in both bands. (D) Frequency response for a configuration with non-PD having higher amplitude in band 1, and PD having higher amplitudes in band 2. CV_NPD is the coefficient of variation for non-PD LFP signal and CV_PD is the coefficient of variation for PD LFP signal.

Figure 4

A snapshot of OFF and ON L-dopa recordings (representing PD and non-PD LFP recordings) of the left DBS lead of dataset A. (A) OFF and ON L-dopa recordings of electrode L0. (B) OFF and ON L-dopa recordings of electrode L1. (C) OFF and ON L-dopa recordings of electrode L2. (D) OFF and ON L-dopa recordings of electrode L3.

Figure 5

A contour plot depicting the effect of increasing/decreasing stimulation frequency on the transition path of a test case (in the XY-location marked “X”) over the feature space. Feature space is that of dataset B.

Figure 6

Surface plot for input-output relationship for: (A) SVM based controller, (B) GMM based controller.

Figure 7

Input-output membership functions for the fuzzy controller driven by SVM state estimates. (A) Membership functions for the state estimates. (B) Membership functions for the rate of change in state. (C) Membership function for the incremental stimulation frequency.

Figure 8

Input-output membership functions for the fuzzy controller driven by GMM state estimates. (A) Membership functions for the state estimates. (B) Membership functions for the rate of change in state. (C) Membership function for the incremental stimulation frequency.

Figure 9

State transition of PD suppression on feature space of patient/dataset E. (A) Showing PD state transition on a feature space using SVM for state estimation, with “X” markers showing start (from PD) and settling (non-PD) positions. The feature space trajectory is indicated in gray. (B) PD state profile for PD suppression using SVM to obtain state estimates. It depicts the modal interval for the non-PD state when SVM is used for state estimation. (C) Showing PD state transition on a feature space using GMM for state estimation, with “X” markers showing start (from PD) and settling (non-PD) positions. The feature space trajectory is indicated in gray. (D) PD state profile for PD suppression using GMM to obtain state estimates. It depicts the modal interval for the non-PD state when GMM is used for state estimation.

Figure 10

Stimulation profile for the state transition shown in Figure 9.

Figure 11

Relative complexity of the critic-actor control driven by GMM and SVM. (A) Normalized complexity for the state estimation stage. (B) Normalized complexity for the fuzzy control stage. Normalized to the maximum for all cases (maximum = 1).