Computational Investigation of Adaptive Deep Brain Stimulation

Christopher Y. Thang and Paul A. Meehan

School of Mechanical & Mining Engineering, The University of Queensland, Brisbane, Australia

Keywords: Closed-loop, Neural Feedback, Adaptive Deep Brain Stimulation, Parkinson’s Disease.

Abstract: Deep Brain Stimulation of the sub-thalamic nucleus (STN) has been proven to be effective at reducing

symptoms of patients with Parkinson’s disease (PD). Currently an implanted pulse generator provides

chronic electrical stimulation to the STN via an electrode and the stimulation parameters are chosen

heuristically. Closed-loop Deep Brain Stimulation (DBS) has been proposed as an improvement to this,

utilising neural signal feedback to select stimulation parameters, adjust the duration of stimulation and

achieve better patient outcomes more efficiently. In this research, potential neural feedback signals were

investigated using a computational simulation of the basal ganglia. It was found that the interspike-interval

in the globus pallidus externus provided a possible metric for ‘on’ and ‘off’ states in Parkinson’s disease.

This parameter was subsequently implemented as neural feedback in an adaptive closed-loop DBS

simulation and was shown to be effective. In particular, the thalamic relaying capability was evaluated using

an Error Index (EI) and the adaptive DBS was found to reduce the EI to 2%, which compared with 20% for

the PD case without DBS. This was achieved using 58% of the stimulation time used during continuous

DBS, indicating a large reduction in DBS energy requirements. This selection and implementation of a

potential neural feedback parameter will assist in developing improved implanted DBS pulse generators.

1 INTRODUCTION

Deep Brain Stimulation (DBS) has proven to be an

effective method for relieving the symptoms of

patients suffering from Parkinson’s disease (PD),

Essential Tremor (ET), dystonia (DT) and other

neurological conditions. During the procedure,

electrodes are inserted into targeted regions of the

basal ganglia (Fig. 1) and connected to an Implanted

Pulse Generator (IPG) positioned subcutaneously

below the clavicle (Coyne, Silburn et al. 2006).

Currently the stimulation parameters for DBS are

chosen heuristically, requiring periodic post-

operative programming sessions to determine the

optimum settings for symptom reduction for 6

months (Marjama-Lyons and Okun 2014). Neural

plasticity effects, progression of the neurological

disease, patient activity states and changes in the

medication may all lead to changing stimulation

requirements over time.

Closed-loop DBS has been proposed as an

alternative to chronic open-loop DBS, utilising

neural feedback signals to regulate the stimulation

parameters. By sensing symptoms and activating

stimulation when it is required, it is anticipated that

the power consumption of the IPG may be reduced,

symptom reduction may be improved and side-

effects minimised. Figure 1 is a diagram of this

closed-loop DBS concept with neural feedback

being used to control the stimulation parameters.

Figure 1: Diagram of closed-loop deep brain stimulation

(Adapted from Huntington's Outreach Project for

Education 2010).

Multiple research groups are currently investigating

the feasibility of a system (Brain Institute, Utah;

Neuromedical Control Systems Lab, John Hopkins

University) and recently closed-loop DBS has been

Y. Thang C. and A. Meehan P..

Computational Investigation of Adaptive Deep Brain Stimulation.

DOI: 10.5220/0005212400660075

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2015), pages 66-75

ISBN: 978-989-758-069-7

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

trialled in patients (Little, FitzGerald et al. 2013).

Prior to this, research has primarily been performed

on primates or computational models. Carron et al.

summarises the current closed-loop control

approaches as delayed and multi-site stimulation;

optimal control; proportional integral derivative

control and adaptive stimulation (Carron, Chaillet et

al. 2013). This work will focus on the latter control

strategy.

A major challenge thus far has been determining

a suitable neural feedback parameter that will

characterise the severity of PD along with

progression for the disease. Initial research has been

conducted, testing the non-markovity spectrum as a

neural measure however this work was focused on

measuring linguistic response (Meehan, Bellette et

al. 2011). It has previously been shown that Local

Field Potential (LFP) recordings from the sub-

thalamic nucleus (STN) show beta oscillations (8-35

Hz) in the PD state (Kühn, Kupsch et al. 2006,

Kühn, Hariz et al. 2008). Recently, Little et al. went

on further to successfully use this parameter as

feedback in a clinical trial of adaptive DBS. Results

showed a reduction in Unified Parkinson’s Disease

Rating Scale (UPDRS) motor scores by 27% in

blinded assessment, when compared with continuous

DBS (Little, FitzGerald et al. 2013). This new

clinical research testing faces many difficulties and

limitations that may be overcome with an efficient,

validated computational model. Hence the present

research aims to investigate and develop adaptive

DBS in a computational environment to better

understand the underlying neural mechanisms

associated with this method of control.

This research focuses on the identification of an

effective feedback parameter for closed-loop DBS

and quantifying the feasibility of using the parameter

in an adaptive DBS feedback system. A

computational model of the basal ganglia based on

the Rubin-Terman model (Rubin & Terman, 2004),

is used to simulate DBS for PD in the STN (So,

Kent et al. 2012). Neural output is then processed

and feedback parameters for PD are investigated. In

particular, the interspike interval is subsequently

implemented as neural feedback in an adaptive

closed-loop DBS simulation and its effectiveness

quantified. It is expected that demonstrating the

feasibility of closed-loop DBS for PD will provide a

basis for future investigations into more efficient and

effective systems for the treatment of PD and other

neurological conditions.

2 METHODS

This research builds upon a well-developed model of

the basal ganglia to develop an adaptive DBS

simulation. The existing basal ganglia model is first

presented, along with simulation parameters. This is

followed by details of two signal analysis methods

used to determine feedback parameters. A third

signal analysis method is also presented which will

be used to characterise the effectiveness of DBS.

2.1 Basal Ganglia Model

The basal ganglia (BG) is involved in the signal

processing of a range of neural functions including

voluntary motor movement, learning, cognition and

emotion. In the BG, information is transmitted

between nuclei via inhibitory and excitatory

projections. Under the canonical model of BG motor

loops, these projections form direct and indirect

pathways through the BG. Figure 2 shows these

pathways and synaptic connections on a cross-

section of the BG.

Figure 2: Diagram of the indirect (blue) and direct (red)

pathways through the basal ganglia with excitatory and

inhibitory connections represented with arrows and flat-

ends respectively (Calabresi, Picconi et al. 2014).

In Figure 2, excitation of the direct pathway from

the cortex has the effect of exciting the putamen,

inhibiting the globus pallidus internus (GPi) and

consequently disinhibition of the thalamus, resulting

in ease of firing. Conversely, excitation of the

indirect pathway from the cerebral cortex results in

excitation of the putamen and consequently

inhibition of the globus pallidus externus (GPe).

This in turn leads to disinhibition of the sub-

ComputationalInvestigationofAdaptiveDeepBrainStimulation

thalamic nucleus (STN), excitation of the globus

pallidus internus (GPi) and inhibition of the

thalamus. When functioning correctly these

competing pathways balance so that the thalamus

operates correctly. A review of indirect/direct model

suggests that it is still the most plausible BG model,

although the possible role of pathway interactions

must be revised in light of recent experimental work

(Calabresi, Picconi et al. 2014).

An existing computational model of the BG,

developed by So et al., was used to simulate neural

output (So, Kent et al. 2012). This model is based on

the Rubin-Terman model (Rubin and Terman, 2004)

and it is a simplified model of the BG with

components of the classical indirect and direct

model. The excitatory and inhibitory synaptic

connections of the computational BG model are

summarised in Figure 3.

Figure 3: Excitatory and inhibitory synaptic connections

used in the Basal Ganglia model.

In this model only the subcortical-thalamic region

(Tha), subthalamic nucleus (STN), globus pallidus

externus (GPe) and globus pallidus internus (GPi)

are modelled. STN neurons are modelled with

excitatory synapses to two GPe and GPi neurons;

GPe neurons inhibit two STN, GPe and GPi neurons

and each GPi neuron inhibits a Tha neuron. The

DBS stimulation is applied to the STN and

sensorimotor excitatory stimuli is inputted to the

Tha.

The individual neuron membrane potentials were

modelled using single-compartment conductance-

based biophysical models of the form (Terman,

Rubin et al. 2002, Rubin and Terman 2004),

m L Na K T Ca ahp in

CIIIIIIII

ab-

=- - - - - - - +

(1)

In (1), C

is the membrane capacitance; V

is the

membrane potential; I

, I

and

ahp

are the

leak, sodium, potassium, low-threshold calcium,

calcium and afterhyperpolarization currents

respectively. Like most conductance based models,

each of these currents is controlled via a time variant

channel activation equation. Depending on the

nuclei, the total cell current is a combination of these

along with any current inputs, I

which may include

DBS, constant bias currents and sensorimotor input.

The complete equations and model parameters used

in this model are available in So et al. Synaptic

currents I

α-β

between neurons are represented as

follows where α is a pre-synaptic and β is a post-

synaptic neuron:

IgVE S

ab ab a ab a 

ëû

(2)

In (2), 

→

is the maximal synaptic conductance

and E

α-β

is the synaptic reversal potential.

The gating

of neurotransmitter receptors, S, varies from 0 to 1

and are described using an average response model.

For STN and GPi efferents, a second order alpha

synapse was used, where u(t) = 1 if the pre-synaptic

cell spikes and is otherwise 0:

0.234 ( ) 0.4 0.04 .

ut z S





(3)

The remaining connections were modelled using

first order alpha synapses of the form:

 



2 1 20 0.04

1/ 1 exp .

SH V S



 

















(4)

In the Parkinson’s disease state, the loss of

dopamine results in disinhibition of the GPi via the

direct pathway and increased inhibition of the GPe,

disinhibition of the STN and further increased

activity of the GPi via the indirect pathway. PD

states were simulated using a reduction in applied

currents to the STN, GPe and GPi as summarised in

Table 1 (So, Kent et al. 2012).

Table 1: Applied currents in PD and Healthy states.

Neural State

app STN



app GPe



app GPi



Healthy

33 µA/cm

20 µA/cm

21 µA/cm

Parkinson’s

23 µA/cm

7 µA/cm

15 µA/cm

This produced firing behaviour consistent with

humans, rodents and primates with PD (So, Kent et

al. 2012). The reduction in the GPe applied current

is the largest which corresponds with disinhibiting

the GPi and STN.

High-frequency DBS input into the STN was

modelled using the Heavyside step function, H,

BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

(sin(2 t)) 1 (sin(2 ( ))) ,

DBS DBS DBS DBS DBS

IiH f H ftppd

éù

=´-+

êú

ëû

(5)

where i

DBS

is the stimulation amplitude, f

DBS

is the

stimulation frequency and δ

DBS

is the pulse width.

Simulations were run in MATLAB (MATLAB 8.3,

The MathWorks Inc., Natick, MA, 2014). The

simulated time was 6000 ms, with 30 neurons per

nuclei, 130 Hz stimulation and a time-step of 0.01

ms using Euler’s forward difference method unless

otherwise indicated.

2.2 Signal Analysis

The interspike interval and synchrony measures are

presented in this section and will be applied to the

spike-trains from the BG model. The third measure,

Error Index, tests for relaying fidelity and will be

used to examine DBS effectiveness.

2.2.1 Interspike Interval

The Interspike Interval (ISI) is defined as the time

between subsequent action potentials of the spike

train. The mean ISI of a neuron is often calculated to

be,

ISI ISI

(6)

where n is the number of spikes and ISI

is the

interval between two spikes. Whilst regularly used

for leaky integrate and fire (LIF) models, it requires

individual spike train data for neurons. This is

suitable for computational models but difficult to

obtain from in-vitro microelectrode recordings

(MER), unless a single-unit MER is used. The

threshold for peak detection was set at -10 mV and

the average and standard deviation of the ISI was

then taken across all 30 neurons, for each nuclei.

This resulted in a dataset of approximately 2,000

spikes for the thalamic region to 14,000 spikes for

the GPi.

2.2.2 Synchrony

It has been proposed that in PD states, increased

synchrony occurs in the GPi and that synchrony is

reduced in the thalamic region (Rubin and Terman

2004). Golomb proposes the following measure for

neuronal synchrony χ (Golomb and Rinzel 1993):

() .

(7)

In (7) the variance of the total voltage is defined as,

() () ,

where ... and ( ) ( ).

Vt Vt

dt V t V t

éù

êú

ëû

(8)

Similarly the variance of individual neurons is

defined as:

() () .

Vt Vts

éù

êú

ëû

(9)

It can be seen that the neuronal synchrony measure

relies on voltage data from a population of neurons

which is only readily obtained from computational

models. For completely asynchronous behaviour, it

is expected that the synchrony measure will decrease

as the number of neurons sampled is increased.

Otherwise, for synchronous and weakly

synchronised neurons, the degree of synchrony will

be constant for varying N.

2.2.3 Error Index

Rubin and Terman have proposed an Error Index

(EI) to characterise the relaying fidelity of the

thalamic region (Rubin and Terman 2004). The EI is

determined as,

()

misses false

EI E E

åå

(10)

where n is the number of input stimuli, E is an error

from either a false positive or a miss. False positives

are defined as spikes occurring without stimulus and

multiple spikes in response to a single stimulus.

Misses are defined as a failure to respond within 10

ms of a stimulus. High EI values correspond with

poor thalamic relaying capability of sensorimotor

input. In this testing, the EI is evaluated over a

sample size of approximately 800 input stimuli. This

EI will be used to characterise the effectiveness of

adaptive DBS in maintaining relaying capacity in the

thalamic region of sensorimotor input and compared

with results for healthy, PD and continuous DBS

states.

3 RESULTS

In this section the signal analysis results are used to

determine an optimum neural feedback parameter.

After being selected, the parameter is then used as

neural feedback in an adaptive DBS simulation. The

effectiveness of the adaptive DBS is compared with

healthy, PD and continuous DBS states, using the

Error Index measure to test for thalamic fidelity.

ComputationalInvestigationofAdaptiveDeepBrainStimulation

3.1 Signal Analysis

Two signal analysis methods are presented here as

possible feedback parameters. The results from the

interspike interval analysis is shown first, followed

by Golomb’s synchrony measure.

3.1.1 Interspike Interval

The Interspike Interval (ISI) results have been

plotted in Figure 4 for each of the nuclei in the

healthy, PD and PD with DBS states. When in the

PD state an increase in ISI can be observed in the

thalamic region due to misfiring. In the STN, the ISI

is reduced when DBS is applied since the neurons

are triggered to fire in unison at the 130 Hz

stimulation frequency. The GPi has a reduced ISI

when DBS is applied and the ISI in the GPe

increases for the PD state.

All four nuclei have an increase in ISI variance

(square of the standard deviation) for the PD states.

In particular, the standard deviation of GPe ISI

increases the most, rising by 380% in the PD state

(see error bars in Figure 4). This compares with a

corresponding 87% increase in ISI. This increase in

variance can be attributed to periodic periods of

spiking in the GPe instead of sustained firing. Once

DBS is applied, GPe ISI and standard deviation of

ISI return to similar values to the healthy state. ISI

variance in the thalamic region also exhibits similar

increases. This could be explained through BG loops

(Figure 3) as reduced activity in the GPe resulting in

disinhibition of the GPi and consequently increased

inhibition of the thalamus. This significant variation

in variance in the GPe shows potential as a possible

neural feedback parameter for ‘on’ and ‘off’ PD

states.

The response of the ISI was then tested for the

four nuclei when the DBS frequency is adjusted.

This was performed to determine if there were any

relationships between DBS frequency and ISI

measures. A batch script was run increasing the

frequency incrementally in intervals of 1 Hz from 0

to 120 Hz. The STN mean ISI was found to be

inversely proportional to frequency, as the STN

neurons are triggered to fire by the DBS.

In Figure 5, high fluctuating values are observed

in the ISI variance for thalamic neurons when

stimulation frequency is less than 52 Hz and

consistent low values are observed above 70 Hz.

From the results in Figure 4, it is assumed that low

variance corresponds with the DBS working

effectively and returning the BG to a healthy state.

In Figure 5, a spike in ISI and ISI variance occurs at

40-50 Hz suggesting that 40-50 Hz DBS may be

counter-effective. Similar trends were observed in

the GP region with stimulation appearing to be

ineffective for frequencies lower than 50 Hz. This is

consistent with the current understanding that low-

frequency DBS (<50 Hz) can be counter-effective

although the neural mechanisms for this are still

unknown (McConnell, So et al. 2012).

Figure 4: Mean Interspike Interval plotted for each of the four nuclei at healthy, Parkinson’s disease and DBS corrected

states with the standard deviation represented as error bars. Stimulation parameters are set at 130 Hz with pulse width of 0.6

ms and amplitude of 350 µA/cm

BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

Figure 5: Thalamic ISI for varying DBS frequency.

The ineffectiveness of low-frequency DBS (<50 Hz)

and consistent effect at high frequencies (>80 Hz)

suggests that DBS frequency would not be an

effective stimulation control parameter.

3.1.2 Synchrony

The synchrony of the four nuclei in the healthy, PD

and PD with 130 Hz DBS states can be seen in

Figure 6. The measure has been plotted as a function

of N, the number of neurons included in the

measure. For completely asynchronous systems, it is

expected that the synchrony measure will go to zero

as N approaches large values. Clearly in this case

this testing is limited since only 30 neurons were

used in each nuclei.

Monotonic decreasing trends towards zero

should be observed if the system is asynchronous

and here it can be seen that the four nuclei are each

not completely asynchronous. The thalamic nuclei

has a 37% reduction in synchrony for the PD state

due to misfiring and GPi inhibition. After DBS is

applied, synchrony increases to value of 1,

indicating that sensorimotor input is being

transferred correctly. In all other regions, the DBS

results in increased degrees of synchrony, with the

greatest increase occurring in the STN where the

DBS is applied.

It appears that synchrony in the thalamic region

could be used as a feedback signal for the PD state,

although the DBS state appears to lead to a higher

degree of synchrony than the healthy state. It should

also be noted that this measure requires the

individual membrane potential data of a large

number of neurons, which is only obtainable in

computer simulations at this stage.

3.2 Adaptive DBS

An adaptive DBS scheme was implemented utilising

ISI variance in the GPe as a trigger for stimulation.

The motivation for this selection was based on the

largest rise between PD and healthy states. The

adaptive DBS stimulation threshold was set for a

GPe ISI standard deviation greater than 13 ms based

on data from the preceding 500 ms. This threshold

was chosen to be slightly above the variance of the

healthy state. A block diagram of this adaptive DBS

system is shown in Figure 7.

Figure 6: Synchrony for the four nuclei plotted as a function of the number of neurons included in the measure.

ComputationalInvestigationofAdaptiveDeepBrainStimulation

Figure 7: Block diagram of the adaptive DBS system.

Dynamic Parkinson’s disease conditions were

created by applying a stepped PD severity input. A

Parkinson’s Severity Factor (PSF) on a scale of 0-1

is used to adjust the PD severity and the applied

current in the STN, GPe and GPi was modified as

follows:

app PD base PD

IIPSFI

=- ´

(11)

The base current values I

base

in (11) were the healthy

applied currents and I

was determined as the

difference between healthy and Parkinsonian states

in Table 1. Once stimulation is triggered, the aDBS

remains on for at least 150 ms before switching off.

The resulting spike trains in the basal ganglia can

be seen in Figure 8 for PD, with adaptive DBS using

GPe ISI variance as a feedback parameter to respond

to stepped PD severity input. For convenience, a

close-up of 1000-3000 ms is shown in Figure 9 so

that the effects of adaptive DBS may be compared

with the PD state. In Figure 8 [A], between 0 and

2000 ms, the BG is in the PD state. By comparing

the neuron spike train (above) with the sensorimotor

input (below), errors in the thalamic relaying

capability can be observed. These errors are

highlighted in Figure 9 [A], with two misfires

observed between 1000-2000 ms. At the same time

PD conditions are evident in Figure 8 [C] & [D]

with bursting firing behaviour in the GPe and GPi.

This is firing behaviour is seen clearly in Figure 9

[C] & [D], (left) with varying intervals between

bursts.

Figure 8: Adaptive DBS (aDBS) using variance of ISI from the GPe as a feedback signal. Between 0-2000 ms the

Parkinson’s severity is at a maximum and aDBS is not turned on; at 2000 ms the aDBS is initiated; between 4000-6000 ms

Parkinson’s severity is stepped down to 50% and at 6000 ms the severity is further reduced to 0%. [A] Spike-train from a

thalamic neuron (above) with sensorimotor input (below); [B] Spike-train from a sub-thalamic nucleus neuron (above) with

DBS stimulation times (below); [C] Spike-train from a globus pallidus externus neuron and [D] Spike-train from a globus

pallidus internus neuron.

BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

When adaptive DBS (aDBS) is initiated at 2000 ms,

the STN stimulation triggers periodically to maintain

consistent firing in the GPe as seen in Figure 8, [B].

Figure 9 shows a close-up of this transition for

thalamic, STN, GPe and GPi neurons. It can be seen

in Figure 9 [B], (right) that the application of

stimulation to the STN triggers high-frequency STN

firing and restores thalamic throughput capability in

Figure 9 [A], (right). The short periods where the

aDBS switches off stimulation does not appear to

lead to errors in the thalamic relaying of

sensorimotor input.

Figure 9: The transition from adaptive DBS being ‘off’ to

‘on’ for an [A] thalamic, [B] sub-thalamic nucleus, [C]

globus pallidus externus and [D] globus pallidus internus

neuron; misfires in the thalamic region are circled in red.

In Figure 9 [C], when aDBS control is initiated the

GPe firing rate is steady during periods when

stimulation is on. Once the standard deviation of the

GPe ISI is reduced below the threshold, the

stimulation switches off and the GPe returns to

bursting behaviour (e.g. 2370-2460 ms). If the

standard deviation of the GPe ISI exceeds the

threshold, stimulation is turned on again and this

‘on’ and ‘off’ cycle can be seen in Figure 8 [B] until

the Parkinson’s severity is lowered. The GPi does

not appear to be affected by the short periods where

the stimulation is switched off which suggests a

delayed response. The continuous firing of the GPi

during these short periods of no stimulation may be

an important factor in the success of the aDBS in

maintaining reliable thalamic throughput.

During 4000-6000 ms in Figure 8, the severity of

Parkinson’s is reduced to 50% and a reduction in

GPe bursting behaviour is observed in [C] due to the

increase in applied current. As a result, the period

between STN stimulation triggering in [B] increases

since there is less variance in the GPe ISI. This

reduction in stimulation ‘on’ time does not appear to

impact the thalamic relaying capability. Finally,

when healthy conditions are imposed from 6000 ms,

regular GPe firing is observed in Figure 8 [C] and

consequently the aDBS control scheme no longer

triggers stimulation. No bursting is observed in the

GPe or GPi and the thalamic neuron transmits the

sensorimotor input correctly.

Figure 10: Error Index results for healthy, Parkinson’s

disease (PD), PD with adaptive DBS and PD with

continuous DBS states.

The Error Index (EI) is used to determine the change

in thalamic relaying capability when adaptive DBS

and continuous DBS are applied. These results are

compared to healthy and PD states in Figure 10 for a

stepped PSF input identical to Figure 11, [A].

Figure 10 shows that adaptive DBS was

successful in reducing the EI from 20% to 2%. The

healthy state had an error index of 1% and the PD

with cDBS state had an EI of 0.2%. Although the

aDBS did not reduce the EI to the levels of cDBS,

occasional misfiring does occur in the healthy state

and this is most likely an acceptable result. The clear

advantage of aDBS here is that only 58% of

stimulation was used in comparison with the cDBS

system whilst achieving a 95% of the reduction

towards the healthy state. If further tuning of the

aDBS is performed along with the implementation

of more advanced control methods, it is possible that

healthy conditions could be achieved with

substantial power savings.

The effectiveness of adaptive DBS was further

quantified by investigating the change in the GPe ISI

standard deviation with respect to Parkinson’s

Severity Factor, as shown in Figure 10. A stepped

PSF input has been inputted to the simulation with

ComputationalInvestigationofAdaptiveDeepBrainStimulation

and without aDBS applied. When no DBS is applied

the standard deviation of the GPe ISI remains well

above the healthy threshold values for a maximum

PSF. It can be noted that the variance has a delayed

response to a stepped reduction to the PSF at 2000

ms. When aDBS is operating, the standard deviation

reaches the threshold regularly and triggers STN

stimulation. Comparing Figure 11 [C] and [D] to [B]

it is deduced that aDBS is effective in maintaining

healthy levels of variance in the GPe spike train.

Figure 11: [A] Stepped inputs of Parkinson’s Severity

Factor over time; [B] Standard deviation of GPe ISI

without aDBS being applied; [C] & [D] Adaptive DBS

with stimulation time (top) and corresponding Std. Dev of

GPe ISI (bottom).

4 CONCLUSIONS

In this investigation we examined the feasibility of

interspike interval (ISI) and synchrony as feedback

parameters for closed-loop DBS. Between the

healthy, Parkinson’s Disease (PD) and PD with DBS

states, an 87% increase in interspike interval was

observed in the globus pallidus externus (GPe). It is

hypothesised that this increase in interval between

spikes results in disinhibition of the globus pallidus

internus (GPi) which in turn inhibits the thalamus

and prevents the thalamus from relaying

sensorimotor input effectively. Underlying

rhythmicity in the GPi appears to further inhibit the

thalamus. In the PD state, the standard deviation of

GPe ISI was found to increase by 380% from the

healthy and PD with DBS states and thus

demonstrated potential as a feedback parameter.

Synchrony in the thalamic region was found to

drop by 37% from the healthy to PD state. It should

be noted that once DBS was initiated, the synchrony

measure in the thalamic region exceeded the healthy

levels. Other regions of the basal ganglia only

exhibited synchrony changes for the PD with DBS

state. Due to the practical difficulties associated with

obtaining multiple neuron recordings in a clinical

setting for the synchrony measure, the standard

deviation of ISI in the GPe was thus investigated

further in a closed-loop DBS simulation.

An adaptive DBS (aDBS) closed-loop control

scheme was used, where stimulation was turned on

or off depending on whether the GPe ISI Std. Dev

exceeded a threshold of 13 ms. The response of the

system was tested for stepped inputs for varying

Parkinson’s Severity Factors (PSF). The aDBS

scheme was successful in improving the thalamic

relaying capability, with an Error Index (EI) of 2%.

This compared with 1% for the healthy state and

20% for PD with no DBS. In this implementation,

the aDBS applied stimulation for 58% of the total

time during a stepped PSF input, indicating a

substantial reduction in DBS power consumption.

Challenges still remain between testing in the

computational environment and implanting this

technology in patients. Although clinical trials have

already successfully been performed by Little et al.

using beta-oscillations as a feedback parameter,

those trials were performed using a wired connection

between dedicated signal processing tools, laptop

and patient. The miniaturisation of these systems

into an IPG has yet to be achieved and this work

faces similar challenges with the signal analysis

tools. The accuracy of ISI interval sampling has also

not been tested in patients for this work although

single-cell microelectrode recordings may offer one

potential solution towards gathering the spike

interval data.

In this computational simulation, variance

(square of the standard deviation) in interspike

interval of the globus pallidus externus has

successfully been used as a feedback parameter for

aDBS. Research into more advanced control

methods such as Proportional-Integral control of

stimulation amplitude may offer further

opportunities to improve stimulation efficiency.

These alternate control methods will require tuning

to overcome the highly non-linear ‘all-or-none’

firing dynamic of neurons. Despite the widespread

use of such controllers in other applications, it is

possible that adaptive DBS may achieve a profile

BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

closer to the healthy state, due to the ‘all-or-none’

nature of the control method.

ACKNOWLEDGEMENTS

The authors would like to acknowledge Professor

Peter Silburn (Asia-Pacific Centre for

Neuromodulation) for his expertise in clinical

aspects of this research along with Professor Tipu

Aziz (Oxford Functional Neurosurgery) for his

research guidance.

REFERENCES

Calabresi, P., B. Picconi, A. Tozzi, V. Ghiglieri and M. Di

Filippo (2014). "Direct and indirect pathways of basal

ganglia: a critical reappraisal." Nat Neurosci 17(8):

1022-1030.

Carron, R., A. Chaillet, A. Filipchuk, W. Pasillas-Lépine

and C. Hammond (2013). "Closing the loop of deep

brain stimulation." Frontiers in Systems Neuroscience

7: 112.

Coyne, T., P. Silburn, R. Cook, P. Silberstein, G. Mellick,

F. Sinclair, G. Fracchia, D. Wasson and P. Stanwell

(2006). "Rapid subthalamic nucleus deep brain

stimulation lead placement utilising CT/MRI fusion,

microelectrode recording and test stimulation." Acta

Neurochirurgica Supplement 99: 49-50.

Golomb, D. and J. Rinzel (1993). "Dynamics of globally

coupled inhibitory neurons with heterogeneity."

Physical Review E 48(6): 4810-4814.

Huntington's Outreach Project for Education, S. U. (2010).

"HOPES Brain Tutorial - Basal Ganglia." Retrieved

14 November, 2014, from http://hopes

.stanford.edu/sites/hopes/files/f_ab18bslgang.gif.

Kühn, A. A., M. I. Hariz, W. Vandenberghe, B. Nuttin, P.

Brown, F. Kempf, C. Brücke, L. Gaynor Doyle, I.

Martinez-Torres, A. Pogosyan, T. Trottenberg, A.

Kupsch, G.-H. Schneider, Neurokirurgi, f. Medicinsk,

n. Farmakologi och klinisk and u. Umeå (2008).

"High-frequency stimulation of the subthalamic

nucleus suppresses oscillatory beta activity in patients

with Parkinson's disease in parallel with improvement

in motor performance." The Journal of Neuroscience

28(24): 6165-6173.

Kühn, A. A., A. Kupsch, G. H. Schneider and P. Brown

(2006). "Reduction in subthalamic 8–35 Hz oscillatory

activity correlates with clinical improvement in

Parkinson's disease." European Journal of

Neuroscience 23(7): 1956-1960.

Little, S., J. FitzGerald, A. L. Green, T. Z. Aziz, P. Brown,

A. Pogosyan, S. Neal, B. Zavala, L. Zrinzo, M. Hariz,

T. Foltynie, P. Limousin and K. Ashkan (2013).

"Adaptive deep brain stimulation in advanced

Parkinson disease." Annals of Neurology 74(3): 449-

457.

Marjama-Lyons, J. and M. Okun (2014). Parkinson's

Disease: Guide to Deep Brain Stimulation Therapy,

National Parkinson Foundation.

McConnell, G. C., R. Q. So, J. D. Hilliard, P. Lopomo and

W. M. Grill (2012). "Effective deep brain stimulation

suppresses low-frequency network oscillations in the

basal ganglia by regularizing neural firing patterns."

The Journal of neuroscience : the official journal of

the Society for Neuroscience 32(45): 15657-15668.

Meehan, P. A., P. A. Bellette, A. P. Bradley, J. E. Castner,

H. J. Chenery, D. A. Copland, J. D. Varghese, T.

Coyne and P. A. Silburn (2011). Investigation of the

non-markovity spectrum as a cognitive processing

measure of deep brain microelectrode recordings.

Biosignals 2011. Rome, Italy, SciTePress: 144-150.

Rubin, J. E. and D. Terman (2004). "High frequency

stimulation of the subthalamic nucleus eliminates

pathological thalamic rhythmicity in a computational

model." Journal of Computational Neuroscience 16(3):

211-235.

So, R. Q., A. R. Kent and W. M. Grill (2012). "Relative

contributions of local cell and passing fiber activation

and silencing to changes in thalamic fidelity during

deep brain stimulation and lesioning: a computational

modeling study." Journal of Computational

Neuroscience 32(3): 499-519.

Terman, D., J. E. Rubin, A. C. Yew and C. J. Wilson

(2002). "Activity patterns in a model for the

subthalamopallidal network of the basal ganglia." The

Journal of Neuroscience 22(7): 2963-2976.

ComputationalInvestigationofAdaptiveDeepBrainStimulation