U.S. patent application number 17/397129 was filed with the patent office on 2021-11-25 for stochastic-switched noise stimulation for identification of input-output brain network dynamics and closed loop control.
The applicant listed for this patent is UNIVERSITY OF SOUTHERN CALIFORNIA. Invention is credited to Maryam Shanechi.
Application Number | 20210361244 17/397129 |
Document ID | / |
Family ID | 1000005767618 |
Filed Date | 2021-11-25 |
United States Patent
Application |
20210361244 |
Kind Code |
A1 |
Shanechi; Maryam |
November 25, 2021 |
STOCHASTIC-SWITCHED NOISE STIMULATION FOR IDENTIFICATION OF
INPUT-OUTPUT BRAIN NETWORK DYNAMICS AND CLOSED LOOP CONTROL
Abstract
Time-efficient identification of a brain network input-output
(IO) dynamics model for brain stimulation includes generating an
input stochastic-switched noise-modulated waveform characterized by
at least one parameter modulated according to a stochastic-switched
noise sequence, inputting the input stochastic-switched
noise-modulated waveform to a clinical brain-response system,
recording one or more time-correlated outputs of the clinical
brain-response system responsive to the input stochastic-switched
noise-modulated waveform, and identifying a brain network IO
dynamics model that optimally correlates the input
stochastic-switched noise-modulated waveform to the one or more
time-delimited outputs of the clinical brain-response system. A
desired brain response to an input electrical signal may be
obtained using the model, such as by modulating the input
electrical signal using a closed-loop control algorithm based on
the brain network IO dynamics model.
Inventors: |
Shanechi; Maryam; (Los
Angeles, CA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
UNIVERSITY OF SOUTHERN CALIFORNIA |
Los Angeles |
CA |
US |
|
|
Family ID: |
1000005767618 |
Appl. No.: |
17/397129 |
Filed: |
August 9, 2021 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
16325297 |
Feb 13, 2019 |
|
|
|
PCT/US2017/046829 |
Aug 14, 2017 |
|
|
|
17397129 |
|
|
|
|
62375385 |
Aug 15, 2016 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
A61N 1/36064 20130101;
G06F 30/20 20200101; A61B 5/7246 20130101; A61B 5/7217 20130101;
A61N 1/0534 20130101; A61N 1/36103 20130101; A61B 5/725 20130101;
A61N 1/36139 20130101 |
International
Class: |
A61B 5/00 20060101
A61B005/00; A61N 1/36 20060101 A61N001/36; G06F 30/20 20060101
G06F030/20; A61N 1/05 20060101 A61N001/05 |
Claims
1. A closed-loop control method for causing a desired brain
response to an input electrical signal, comprising: providing an
input electrical signal to a true brain system via at least one
input electrode; detecting a response of the true brain system to
the input electrical signal via at least one output electrode;
comparing the response to a desired brain response; and modulating
the input electrical signal, according to a closed-loop control
algorithm based on a brain network IO dynamics model for brain
stimulation developed by correlating a stochastic-switched
noise-modulated waveform input to the true brain system output,
until the response matches the desired brain response.
2. The method of claim 1, wherein the brain network IO dynamics
model optimally correlates the input stochastic-switched
noise-modulated waveform to the true brain system output.
3. The method of claim 1, wherein the desired brain response
mitigates a neuropathic condition.
4. The method of claim 1, wherein the desired brain response
heightens a brain performance parameter.
5. An apparatus for causing a desired brain response to an input
electrical signal, comprising a processor coupled to a memory and
to a waveform generator, the memory holding instructions that when
executed by the processor cause the apparatus to perform: providing
an input electrical signal to a true brain system via at least one
input electrode; detecting a response of the true brain system to
the input electrical signal via at least one output electrode;
comparing the response to a desired brain response; and modulating
the input electrical signal, according to a closed-loop control
algorithm based on a brain network IO dynamics model for brain
stimulation developed by correlating a stochastic-switched
noise-modulated waveform input to the true brain system output,
until the response matches the desired brain response.
6. The apparatus of claim 5, wherein the memory holds further
instructions for identifying the brain network IO dynamics model by
optimally correlating the input stochastic-switched noise-modulated
waveform to the true brain system output.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This application is a divisional of U.S. patent application
Ser. No. 16/325,297, filed Feb. 13, 2019, which is a 371 national
filing of International Patent Application No. PCT/US2017/046829,
filed Aug. 14, 2017, which claims the benefit and priority of U.S.
provisional patent application Ser. No. 62/375,385, filed Aug. 15,
2016. Each of the aforementioned applications is incorporated
herein by reference in its entirety.
FIELD
[0002] The present disclosure relates to identification of
input-output (1/O) brain network dynamics, and related applications
in neuroscience and therapy, for example, deep brain stimulation
for treatment of neurological disorders.
BACKGROUND
[0003] Millions of patients suffer from neurological disorders such
as Parkinson's disease (PD), epilepsy and major depression. For
patients who are refractory to traditional treatments such as
medication, deep brain stimulation (DBS) provides an alternative
treatment. DBS typically works by applying a periodic electrical
pulse train stimulation to an electrode placed in an appropriate
region of the brain, e.g., the subthalamic nucleus, to treat a
certain neurological disorder, e.g., PD. DBS has been shown to
effectively alleviate symptoms of advanced PD, refractory epilepsy,
major depression, obsessive-compulsive disorder, and dystonia.
Despite the clinical benefits of DBS, its widespread application is
currently limited by its partial efficacy and potential side
effects. For example, in PD, DBS treatments are commonly conducted
in an open-loop and trial-and-error-based manner, whereby a
clinician selects the appropriate DBS parameters such as frequency,
amplitude and pulse width. It can take months for an expert to find
initial effective DBS parameters via trial-and-error. Moreover, DBS
parameters are only adjusted every few months to maximize clinical
improvements while minimizing stimulation-induced side effects.
However, symptom dynamics of PD can be much faster, which might
have caused DBS to leave some clinical fluctuations partly
uncontrolled.
[0004] Therefore, there has been extensive research on developing
closed-loop or adaptive DBS systems for better treatment of PD and
epilepsy. These studies suppose that neural signals such as local
field potential (LFP) and neural spikings could represent the
dynamics of the disorders and can be used as feedback to adjust the
DBS treatment in closed loop. However, current closed-loop DBS
systems for PD and epilepsy have only used simple on-off control
schemes. These closed-loop DBS systems have two limitations. First,
they only control the timing of the DBS pulse trains, i.e., when to
start DBS and when to stop, but do not the pulse train amplitude
and frequency during stimulation. For example, one study used an
on-off control triggered by a threshold of LFP amplitude, but the
effective DBS amplitude and frequency were still determined by
experts in open loop. Second, even these simple on-off controllers
have not been fully automated, e.g., determining the LFP amplitude
threshold has been done by an expert via experience, without
automatic optimization of treatment effect. While prototype
closed-loop DBS systems have been shown to be more effective than
standard open-loop DBS in both human clinical experiments and
animal models, the above two limitations need to be addressed to
allow for optimized wide-range clinical applications of closed-loop
DBS systems. In addition, a principled system-identification
framework capable of providing optimal or improved input-output
(IO) dynamics estimation is currently lacking.
[0005] It would be desirable, therefore, to provide a
time-efficient method for defining a model and model parameters
capable of predicting the response of particular brain system to an
input electrical signal and related therapeutic protocols that
overcome these and other limitations of the prior art.
SUMMARY
[0006] This summary and the following detailed description should
be interpreted as complementary parts of an integrated disclosure,
which parts may include redundant subject matter and/or
supplemental subject matter.
[0007] A common aspect of prior closed-loop DBS systems is a lack
of model-based design. A "model-based design" of closed-loop DBS
system or brain-machine interface (BMI) means a design that uses a
computational model to quantify input-output (IO) dynamics of
neural feedback signals in response to DBS parameters, and a
closed-loop controller designed based on this model that optimizes
the treatment effect. FIG. 1A shows a system 100 using a brain
network IO dynamical model 102, output monitoring of neural
activity y.sub.t, and a closed loop controller 104 that generates
electrical stimulation u.sub.t to achieve a specific therapeutic
purpose based on a predictive brain model and real-time input
y.sub.t. Modeling and identification of network level I/O dynamics
is essential for devising principled, high-performance closed-loop
DBS controllers. The present disclosure teaches a general
computational framework to identify brain network I/O dynamics and
design model-based closed-loop DBS systems or brain-machine
interfaces (BMIs) for treatment of a wide range of neurological
disorders, or other applications.
[0008] Up until now, model-based closed-loop DBS systems have only
been studied in simulations. For example, for treatment of tremor,
a prior researcher built a linear autoregressive model to relate
the input DBS amplitude to the output LFP waveform, and based on
this model designed a minimum variance controller to regulate the
LFP power spectrum to a reference profile derived under tremor free
conditions. Simulation studies show the possibility of model-based
controllers to control the DBS frequencies or amplitudes
continuously to target desired neural firing rate pattern or LFP
spectrum that represent a healthy state. However, one important
topic that prior work has omitted is estimation of the network IO
model parameters, or system-identification. The present inventors'
prior work assumed perfect knowledge of the IO model parameters, or
consisted of heuristic IO model identification without optimizing
system-identification accuracy. Accurate identification of the IO
dynamics has crucial influence on the performance of closed-loop
controllers, and can be achieved using systems and methods as
taught herein.
[0009] Closed-loop deep brain stimulation (DBS) systems can improve
traditional open-loop DBS treatment of neurological disorders such
as Parkinson's disease and depression by adjusting DBS parameters
in real time based on neural feedback. However, current closed-loop
DBS systems have not been model-based, and only used simple on-off
control strategies that may not have optimal treatment effects. The
fundamental limitation is that current systems lack appropriate
quantification of the input-output (IO) dynamics from input DBS
parameters to output neural feedback signals. Modeling and
identification of such network-level IO dynamics is essential for
devising principled, high-performance closed-loop DBS
controllers.
[0010] The present disclosure concerns applications using a general
computational framework to identify brain network IO dynamics and
design model-based closed-loop DBS systems or brain-machine
interfaces (BMIs) for treatment of a wide range of neurological
disorders. The approach uses linear state-space models to describe
brain network IO dynamics and design open-loop
system-identification experiments to estimate the model parameters.
A novel open-loop input DBS waveform-a pulse train modulated by
binary noise (BN) DBS parameters-satisfies both optimality
constraints for system-identification and safety constraints for
clinical operation. We further extend the BN-modulated DBS waveform
to a generalized BN (GBN)-modulated DBS waveform to reduce the
identification time.
[0011] The disclosure further describes a clinical closed-loop
simulation testbed with the Neuro Omega.TM. microelectrode
recording and stimulation system. Using this realistic closed-loop
testbed, the identification framework can be tested in closed-loop
control simulations, for example, by observing mood symptom control
in depression or other condition. Our simulation results show that
BN-modulated DBS waveform enabled accurate IO dynamics
identification, and that a closed-loop controller designed from
BN-identified models can achieve tight control of mood symptoms.
Steady-state and transient performance is not significantly
different from a best unrealizable controller that can be derived
from a true brain network model. The closed-loop control
performance is robust to stochastic noises, unknown disturbances
and stimulation artifacts. Finally, when identification time is
limited, performance of BN-modulated DBS waveform can be further
improved by GBN modulated DBS waveform by incorporating
time-constant information of the underlying brain network. Our
results have significant implications for developing future
closed-loop DBS treatments of a wide range of neurological
disorders.
[0012] In an aspect of the disclosure, a method for time-efficient
identification of a brain network input-output dynamics model for
brain stimulation includes generating an input stochastic-switched
noise-modulated waveform characterized by at least one parameter
modulated according to a stochastic-switched noise sequence. The
stochastic-switched noise-modulated waveform may be, or may
include, a binary waveform. The at least one parameter may be
selected from frequency, amplitude, or pulse width. In an aspect,
the stochastic-switched noise-modulated waveform may be generalized
by a tunable parameter based on a ratio of a minimum switching time
(T.sub.sw) to a time constant (T.sub.c) for the brain network IO
model. For example, the tunable parameter is determined by
1-T.sub.sw/T.sub.c. In another aspect, the method may include
estimating a value of T.sub.c by applying an on-off input pattern
to the clinical brain-response system in open loop mode and
measuring a rise time for achieving a defined initial portion of a
transition period. The defined initial portion may be defined, for
example by 1/e. The estimating may further include repeating the
applying and the measuring thereby obtaining multiple values of the
rise time, and averaging the multiple values.
[0013] The method may further include inputting the input
stochastic-switched noise-modulated waveform to a clinical
brain-response system. The brain-response system may be an
electronic system that simulates brain response with
microelectrodes. In an alternative, or in addition, the
brain-response system may be an electronic system that provides and
optionally records an electrical response of an actual brain using
microelectrodes or equivalent monitoring tools. For example, the
brain-response system may include an electrode implanted in a
living brain with the outputs measured by an electrode array.
[0014] The method may further include recording one or more
time-correlated outputs of the clinical brain-response system
responsive to the input stochastic-switched noise-modulated
waveform. The method may further include identifying, as a brain
network IO dynamics model for brain stimulation, a mathematical
model that optimally correlates the input stochastic-switched
noise-modulated waveform to the one or more time-delimited outputs
of the clinical brain-response system. In an aspect, the
mathematical model is a linear state-space model.
[0015] In another aspect of the disclosure, a method for causing a
desired brain response to an input electrical signal includes
providing an input electrical signal to a true brain system via at
least one input electrode, detecting a response of the true brain
system to the input electrical signal via at least one output
electrode, comparing the response to a desired brain response; and
modulating the input electrical signal, according to a closed-loop
control algorithm based on a brain network IO dynamics model for
brain stimulation developed by correlating a stochastic-switched
noise-modulated waveform input to the true brain system output,
until the response matches the desired brain response. In an
aspect, the brain network IO dynamics model optimally correlates
the input stochastic-switched noise-modulated waveform to the true
brain system output. In other aspects, the desired brain response
may mitigate a neuropathic condition, or heighten a brain
performance parameter. The brain network IO dynamics model may be
developed as described in more detail elsewhere herein.
[0016] In related aspects, an apparatus for any of the
aforementioned purposes may include a hardware processor coupled to
a memory and to a waveform generator, the memory holding
instructions that when executed by the processor cause the
apparatus to perform the operations of the methods as summarized
above.
[0017] To the accomplishment of the foregoing and related ends, one
or more examples comprise the features hereinafter particularly
pointed out in the claims and fully described in the detailed
description following the brief description of the drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
[0018] The features, nature, and advantages of the present
disclosure will become more apparent from the detailed description
set forth below when taken in conjunction with the drawings, in
which like element numerals are used to indicate like elements.
[0019] FIG. 1A is a block diagram showing components of a
model-based closed loop controller for DBS.
[0020] FIG. 1B is a block diagram showing components of a multiple
input multiple output (MIMO) linear state space model (LSSM) for
use in DBS closed loop control.
[0021] FIG. 1C is a diagram of a method for open-loop
system-identification used for estimating parameters of an LSSM
model.
[0022] FIG. 2A is a diagram illustrating an example of a random
binary noise (BN) input pattern for generating a DBS waveform.
[0023] FIG. 2B is a diagram illustrating an example of a biphasic
BN pulse train with two adjustable parameters for use as a DBS
waveform.
[0024] FIG. 2C is a diagram illustrating an example of a
generalized binary noise (GBN) input pattern wherein the
probability of frequency change is adjustable, for generating a DBS
waveform.
[0025] FIG. 2D is a diagram illustrating an example of
GBN-modulated pulse train for use as a DBS waveform.
[0026] FIG. 3A is a block diagram illustrating components of a
brain-simulation testbed.
[0027] FIG. 3B is a functional block diagram illustrating
operational aspects of simulated ECoG generation by a computational
component of the testbed.
[0028] FIG. 3C is a functional block diagram illustrating
operational aspects of generating DBS parameters by another
computational component of the testbed.
[0029] FIG. 3D is a flow diagram illustrating an overview of a
design process for a closed-loop controller.
[0030] FIG. 4A is a power spectral density (PSD) graph illustrating
PSD for four different types of input patterns.
[0031] FIG. 4B shows four related Bode plots each for a different
input pattern.
[0032] FIG. 4C shows boxplots comparing IO dynamics estimation
error for different input patterns.
[0033] FIG. 4D shows boxplots of the noise dynamics estimation
errors for different input patterns across 100 random underlying
LSSMs.
[0034] FIG. 5A shows a chart of a binary noise (BN) controller in a
closed-loop Monte Carlo simulation.
[0035] FIG. 5B shows a chart of a rectangular (rect) controller in
a closed-loop Monte Carlo simulation.
[0036] FIG. 5C shows a chart of a single-sine controller in a
closed-loop Monte Carlo simulation.
[0037] FIG. 5D shows a chart of a multi-sine controller in a
closed-loop Monte Carlo simulation.
[0038] FIG. 5E shows controller comparisons for a sample Simulated
Brain Network across 10 trials.
[0039] FIG. 5F shows controller comparisons in Monte-Carlo
experiments over 100 Simulated Brain Networks.
[0040] FIG. 6A is a chart comparing BN controllers with an oracle
controller and target in a Neuro Omega.TM. simulation, for a first
brain network.
[0041] FIG. 6B shows boxplots comparing results for different
measures across 10 trials of the comparison represented by FIG.
6A.
[0042] FIGS. 6C-D parallel FIGS. 6A-B, for a second brain
network.
[0043] FIG. 7A shows the GBN pattern for the GBN controllers in a
Neuro Omega.TM. simulation with a time-constrained scenario, for a
first brain network.
[0044] FIG. 7B shows boxplots comparing results for different
measures across 10 trials of GBN controller in FIG. 7A with a BN
controller and an oracle controller and target in a Neuro Omega.TM.
simulation with a time-constrained scenario, for a first brain
network.
[0045] FIGS. 7C-D parallel FIGS. 7A-B, for a second brain
network.
[0046] FIGS. 8A-F are charts comparing three different scenarios in
two different simulated brain networks, showing robustness of a BN
controller to unknown mood disturbance levels.
[0047] FIG. 9 is a flowchart illustrating aspects of a method for
time-efficient identification of a brain network input-output (IO)
dynamics model for brain stimulation such as deep brain
stimulation.
[0048] FIG. 10 is a block diagram illustrating aspects of an
apparatus for performing the method shown in FIG. 9.
[0049] FIG. 11 is a flowchart illustrating aspects of a method for
treatment of a neurological disorder using a closed-loop controller
of stimulation constructed based on the noise-modulated
waveform.
[0050] FIG. 12 is a block diagram illustrating aspects of an
apparatus for performing the method shown in FIG. 11.
DETAILED DESCRIPTION
[0051] A linear state-space model is used to describe brain network
IO dynamics and design open-loop system-identification experiments
to estimate the model parameters. A novel open loop input brain
stimulation waveform--e.g., a pulse train modulated by binary noise
(BN) DBS parameters--satisfies both optimality constraints for
system-identification and safety constraints for clinical
operation. The BN-modulated brain stimulation waveform is
generalized to a generalized BN (GBN)-modulated stimulation
waveform, reducing identification time and improving time
efficiency. The brain stimulation waveform may be for deep brain
stimulation (DBS).
[0052] The GBN is used as input to a clinical closed-loop
simulation testbed with the Neuro Omega.TM. microelectrode
recording and stimulation system. In a therapeutic application, the
testbed is replaced by a true brain system with electrode inputs
and outputs. Using this realistic closed-loop testbed or true brain
system, parameters of the model for system operation are
identified, for example, in an identification framework to build a
closed-loop controller for mood symptom control in depression.
After the closed loop controller is identified, closed loop control
for the targeted brain state can be based on the identified BN
waveform.
[0053] Simulation results show that the BN-modulated DBS waveform
enabled accurate IO dynamics identification. A closed-loop
controller designed from BN-identified models should be capable of
achieving tight control of mood symptoms. Its steady-state and
transient performance should be close to a best unrealizable
controller that is derived assuming knowledge of a true brain
network model. The closed-loop control performance of the model
proved robust to stochastic noises, unknown disturbances and
stimulation artifacts. When identification time is limited,
performance of BN-modulated DBS waveform can be further improved by
use of the GBN-modulated DBS waveform via incorporating the
time-constant information of the underlying brain network. The
results have significant implications for developing closed-loop
DBS treatments for a wide range of neurological disorders, or other
applications.
[0054] Design of the methods and apparatus as described herein may
be based on a computational framework for identification of brain
network IO dynamics, with the purpose of designing model-based
DBS/stimulation controllers. A multiple-input-multiple-output
(MIMO) linear state-space model (LSSM) is described for modeling
the IO dynamics because LSSM (i) allows for efficient
identification from possibly high-dimensional neural recordings,
and (ii) is amenable to real-time and robust feedback-control. More
detailed justifications of using LSSM are discussed later in the
specification. With the model developed, LSSM parameters are
identified by using open-loop system-identification experiments
applying an appropriate input DBS waveform to stimulate the brain,
and then collecting the input-output data to estimate LSSM
parameters.
[0055] A crucial component of system identification is the choice
of the input DBS waveform, which will critically influence the
accuracy of model identification, and subsequently of closed-loop
control. Therefore, optimal DBS input signals should be designed
and applied for brain network IO identification. Current input DBS
waveforms for brain network IO identification have only been chosen
heuristically, without theoretical guarantees for optimality. For
brain network IO identification, an optimal input design should at
least satisfy the following constraints: [0056] (i) To ensure
accurate IO identification, input waveforms should sufficiently
excite the network neural activity so that maximal information is
obtained from the IO data. For estimating LSSMs, this necessitates
an input with white spectrum. [0057] (ii) To ensure clinical
safety, input DBS waveforms should be constrained to take the form
of pulse trains that are charge-balanced in short time intervals.
[0058] (iii) An optimal input DBS waveform should be easy to
implement in the DBS device. [0059] (iv) The input DBS waveform
should allow for a short identification time since in typical
clinical scenarios the time available for stimulation to perform
system-identification is limited.
[0060] These constraints pose a challenging DBS/stimulation input
design problem. The spectrum of traditional DBS pulse train
waveforms is far from white. Even though pulse trains with Poisson
arrivals (which are approximately "white") have been used to study
brain's LFP under optogenetic stimulation, there exists no evidence
that such waveforms are clinically safe and effective in electrical
DBS. Accordingly, novel aspects of the present disclosure include a
design for input DBS waveforms to solve the above constrained
design problem for IO brain network identification with a LSSM
model structure. The design provides a novel input DBS waveform
that satisfies the optimal spectrum requirement, is clinically safe
and is easy to implement. This DBS waveform is constructed as a
pulse train whose parameters such as amplitude and frequency are
stochastically changed according to a binary noise (BN) sequence.
As used herein, this input DBS waveform may be referred to as a
"BN-modulated" DBS waveform. In addition, to reduce system
identification time, the BN-modulated DBS waveform may be extended
to a generalized BN (GBN)-modulated DBS waveform by further
estimating (using on-off stimulation) and incorporating in the
waveform design the time-constant of the underlying brain network.
This further optimizes BN to reduce identification time.
[0061] In addition, developing closed-loop DBS systems would
significantly benefit from extensive and realistic simulations to
verify performance and robustness before human experiments.
However, so far, simulation studies for closed-loop DBS have been
limited to running Monte Carlo experiments in a computer program,
which makes modeling factors such as delays, recording noises,
disturbances, and stimulation artifacts difficult. Accordingly, the
present disclosure includes a clinical closed-loop simulation
testbed within which modeling and closed-loop control strategies
can be validated using the same recording and stimulation hardware
that would be used in human experiments. The clinical real-time
closed-loop simulation testbed uses the Neuro Omega.TM.
microelectrode recording and stimulation system (Alpha Omega Co.
USA Inc., Alpharetta, Ga.), and incorporates stochastic noises,
disturbances, time-delays, and recording stimulation artifacts that
are difficult to mimic purely using a computer program. This
testbed fully resembles a clinical scenario except that the neural
signal is generated by a simulated brain network. The testbed
enables testing of both open-loop system-identification and
closed-loop control.
[0062] Combined, the novel features summarized above allow testing
of a novel modeling and identification framework for brain IO
dynamics within a realistic clinical testbed using two sets of
experiments. First, the framework is tested for open-loop
system-identification by implementing the BN-modulated DBS
waveform. It can be demonstrated thereby that the new input DBS
waveform is critical for accurate estimation of the brain network
IO model. In addition, the testbed enables accurate estimation of
the time-scale of the underlying network, enabling in turn design
of the new GBN-modulated DBS waveform to significantly reduce the
identification time. Second, the framework can be tested by
implementing it within the clinical testbed using a stochastically
optimal linear-quadratic-Gaussian (LQG) closed-loop controller
designed from the identified LSSM. Taking mood symptom control in
depression as an example, it can be shown that the proposed
modeling and identification framework using the BN-modulated DBS
waveform is critical to achieve accurate closed-loop control of
mood symptoms.
[0063] Network IO Dynamical model and System Identification. FIG.
1B shows aspects of a discrete time-invariant MIMO LSSM 110 to
describe brain network IO dynamics for a hidden brain state 114.
The MIMO LSSM may be based on a system of linear equations, for
example:
x.sub.t+1=Ax.sub.t+Bu.sub.t+w.sub.t
y.sub.t=Cx.sub.t+Du.sub.t+v.sub.t (1)
[0064] In the system (1) and LSSM 110, u.sub.t .di-elect
cons.R.sup.m represents the input electrical stimulation, y.sub.t
.di-elect cons.R.sup.k represents the observed neural activity, and
x.sub.t .di-elect cons.R.sup.n represents the hidden brain state
114. w.sub.t .di-elect cons.R.sup.n and v.sub.t .di-elect
cons.R.sup.k are white Gaussian noises with zero mean to summarize
modeling errors and unmeasured disturbances. The LSSM parameters
are the model order n, the matrices A .di-elect
cons.R.sup.n.times.n, B .di-elect cons.R.sup.n.times.m, C .di-elect
cons.R.sup.k.times.n and D .di-elect cons.R.sup.k.times.m, and the
noise covariance defined as
E .function. [ ( w i v j ) .times. ( w i T .times. v j T ) ] = [ Q
S S T R ] .times. .delta. i .times. j , ##EQU00001##
where Q .di-elect cons.R.sup.n.times.n, R.di-elect
cons.R.sup.k.times.k, S .di-elect cons.R.sup.n.times.k, and
.delta..sub.ij=1 if i=j and 0 otherwise. The DBS parameters are
used by the channel modulators 112 to producer multiple channels of
the input electrical stimulation.
[0065] Pulse amplitude and frequency are important DBS parameters
that can influence treatment outcomes of PD in open-loop simulation
and clinical studies. Thus, the input u.sub.t may be defined as the
DBS parameters, i.e., pulse frequency and amplitude, at each
stimulation site for the brain. With m.sub.1 stimulation sites,
u.sub.t may have dimension m=2m.sub.1. The input DBS waveform may
be a continuous square pulse train 210 modulated by these DBS
parameters, for example, as shown by FIG. 2B. As used herein, the
DBS parameters u.sub.t may be referred to as an "input pattern",
and the stimulation pulse train as "DBS waveform". The observed
neural activity y.sub.t may be defined as the signal features
associated with the neural recordings. Signal features may include,
for example, firing rates of neural spikes, power spectrum of
electrocorticography (ECoG), phase-amplitude-coupling of LFP, or
any other useful feature. The symbol k denotes the total number of
neural features used for model development and input control.
[0066] LSSM parameters may be estimated by developing an open-loop
system-identification approach 120 diagrammed in FIG. 1C.
Critically, this system-identification requires the design of novel
input patterns {u.sub.t}.sub.t.sup.T=1 (where T is the duration of
stimulation) to generate an informative open-loop DBS waveform. An
open-loop IO dataset 122 .sup.T={u.sub.t, y.sub.t}.sub.t.sup.T=1 is
collected based on the DBS waveform. The open-loop
system-identification also requires the use of an appropriate
system-identification algorithm to estimate the LSSM parameters
from the IO dataset .sup.T. The LSSM parameters 127 are estimated
using subspace algorithms 126 based on the LSSM model structure 124
wherein Akaike's information criterion (AIC) may be used to select
the optimal model order. While subspace methods have been
extensively used in engineering and industry, the methods are
disclosed here for the first time for identifying brain network IO
dynamics. Designs of optimal input DBS waveforms for open-loop
identification of brain network IO dynamics are disclosed in the
following two sections.
[0067] BN-Modulated DBS Waveform In Open-Loop
System-Identification. Seemingly contradictory requirements for
optimal input design, as pointed out above, make the input design
problem challenging. The challenge may be overcome by separately
designing a theoretically optimal input pattern and a
clinically-safe DBS waveform. First, for the LSSM, u.sub.t is
defined as the DBS parameters, hence the optimality for
system-identification requires the input pattern
{u.sub.t}.sub.t.sup.T=1 to have an approximate white spectrum only
in the DBS parameter domain. Second, clinical safety further
requires the DBS waveform to be charge-balanced in short time
intervals, therefore, the DBS waveform should consist of
charge-balanced pulses. Third, while there are many input patterns
that have white spectrum, they may not be easy to implement for DBS
devices. Considering these three facts, we construct a novel
DBS/stimulation waveform by choosing the input pattern (or
equivalently, the DBS parameters) as binary noise (BN). As we
describe in detail below, with this choice, our DBS waveform
satisfies all constraints: (i) it is white in the DBS parameter
space; (ii) it is clinically safe as it consists of change-balanced
pulses; (iii) it is easy to implement as it only requires switching
the DBS parameters between two values.
[0068] BN sequences may be used to design the input pattern. A BN
sequence is a stochastic signal 200 that randomly switches between
two fixed signal levels as shown in FIG. 2A. The BN sequence 200
has approximately white spectrum. A BN sequence is generated
stochastically; at each time point tT.sub.sw (t=1, 2, 3, . . . ),
where T.sub.sw is the basic (fastest) switching time, the signal
either remains at the previous level L.sub.1
(u.sub.t=u.sub.t-1=L.sub.1) or changes to the other available level
L.sub.2 (u.sub.t=L.sub.2 .noteq.u.sub.t-1) with equal
probability,
Pr(u.sub.t=u.sub.t-1)=0.5;Pr(u.sub.t.noteq.u.sub.t-1)=0.5. (2)
[0069] The above BN sequence 200 has a spectrum that is white: its
asymptotic discrete-time frequency spectrum .PHI.(.omega.) is given
by
.PHI. .function. ( .omega. ) = T sw , .A-inverted. .omega.
.di-elect cons. 0 , .pi. T s .times. w . ##EQU00002##
It should be noted that the frequency .omega. here means how
frequently the DBS parameters are changing overtime, and should not
be confused with the pulse frequency in the DBS waveform. .omega.=0
means the DBS parameters are constant over time. BN is generally
regarded as an "optimal" test signal in the system-identification
literature since it excites all frequencies .omega..di-elect
cons.0,
.pi. T s .times. w ##EQU00003##
with approximately equal power.
[0070] To design the DBS waveform, independent BN sequences may be
applied to each input dimension in u.sub.t, i.e., to each DBS
parameter. Thus, the DBS waveform is a pulse train consisting of
pulses modulated by BN DBS parameters, termed as BN-modulated DBS
waveform. As an example, consider a biphasic DBS pulse train 210
with two adjustable parameters, i.e., pulse frequency and
amplitude, as shown in FIG. 2B. After choosing a proper T.sub.sw,
two independent BN sequences BN.sub.t.sup.1 and BN.sub.t.sup.2 are
generated, with time-varying frequency f.sub.t (Hz) and amplitude
at (mA) set as f.sub.t=BN.sub.t.sup.1, a.sub.t=BN.sub.t.sup.2.
Here, the two levels of BN represent a range of the time-varying
DBS parameters. In this example of a biphasic pulse train 210,
pulse frequency, L.sub.1=50, L.sub.2=100; and pulse amplitude,
L.sub.1=3, L.sub.2=6 as diagrammed in FIG. 2B. Therefore, during
the time interval [t.sub.0T.sub.sw,t.sub.0T.sub.sw+T.sub.sw], the
BN-modulated DBS waveform consists of biphasic pulse trains with
pulse frequency BN.sub.t.sub.0.sup.1 and amplitude
BN.sub.t.sub.0.sup.2.
[0071] GBN-Modulated DBS Waveform And Identification Time. A BN
input pattern is theoretically optimal for system-identification
assuming no constraint on the required identification time T in
.sup.T. However, there exist cases where the time available for
stimulation to perform system-identification is limited. In this
time-limited case, the performance of BN input pattern in
system-identification will degrade, as explained below under the
heading "GBN-Modulated DBS Waveform Reduced System-Identification
Time." Hence improvement of the BN input pattern may be desirable
for time-limited cases. Improvement may be achieved by
incorporating the time-constant information of the underlying brain
network dynamics to reduce T.
[0072] The idea is to utilize the concept of generalized binary
noise (GBN). GBN is a direct generalization of BN, where u.sub.t
stays at the same level with probability .rho. rather than with
fixed probability 0.5, i.e.,
Pr(u.sub.t=u.sub.t-1)=p;Pr(u.sub.t.noteq.u.sub.t-1)=1-p. (3)
[0073] Compared with BN, which has a fixed p=0.5, GBN has an extra
parameter p that can be tuned. Instead of the white spectrum of BN,
a GBN with p>0.5 has more low frequency components and vice
versa for p<0.5. An example of a GBN signal 220 with p=0.9 is
diagrammed in FIG. 2C. Hence p can be tuned to achieve more
accurate estimation of frequency ranges that are important in the
brain network dynamics by putting more weight on these ranges. The
time-constant T.sub.c of the brain network dynamics guides the
choice of p as
p = 1 - T s .times. w T c . ( 4 ) ##EQU00004##
The fastest switching time T.sub.sw may be selected to be less than
T.sub.c, such that BN has white spectrum over frequency
interval
.omega. .di-elect cons. [ 0 , .pi. T s .times. w ] ##EQU00005##
that covers the main frequency components
[ 0 , .pi. T c ] ##EQU00006##
of the brain network dynamics.
[0074] While it is possible to use the time-constant information to
instead tune the basic switching period of the BN input pattern
(e.g., a BN input pattern with
T sw ' = T c 2 , ##EQU00007##
which keeps p=0.5), for brain network dynamics with
T.sub.c>>T.sub.sw, this tuned BN input pattern has the
disadvantage that its spectrum vanishes at
.omega. = 2 .times. k .times. .pi. T s .times. w , ##EQU00008##
k=1, 2, 3, . . . . This introduces a considerable number
( around .times. .times. T c T s .times. w ) ##EQU00009##
of undesirable "gaps" within the frequency band
[ 0 , .pi. T s .times. w ] . ##EQU00010##
Therefore this tuned BN does not have white spectrum across the
entire band
[ 0 , .pi. T s .times. w ] ; ##EQU00011##
it cannot excite the frequency components at these "gap
frequencies", and can lead to a considerable loss of information.
In contrast, GBN spectrum does not vanish at any point within
[ 0 , .pi. T s .times. w ] . ##EQU00012##
Hence under the condition that the identification time is the same,
GBN input pattern can outperform BN input pattern in identifying
linear dynamics as long as the time-constant information of the
dynamics is available.
[0075] Therefore, to design the GBN input pattern, a rough estimate
of the time-constant T.sub.c is needed. This estimate may be
obtained by conducting an extra open-loop experiment where an
on-off input pattern is applied, the rise time needed for each
output feature to finish 36.79%, i.e.,
1 e , ##EQU00013##
of the entire transition period is calculated, and these rise times
across all output features are averaged to get the estimated
time-constant {circumflex over (T)}.sub.c. The duration of this
extra stimulation is typically much shorter than the extra
identification time BN needs, as explained in more detail below
under the heading "GBN-Modulated DBS Waveform Reduced
System-Identification Time."
[0076] Closed-Loop Simulation Testbed With Neuro Omega.TM.. A new
real-time closed-loop simulation testbed 300 is needed to validate
the brain network IO identification framework, for example, as
shown in FIG. 3A. A suitable testbed may be constructed using the
clinical Neuro Omega.TM. microelectrode recording and stimulation
system 310 (Alpha Omega Co. USA Inc., Alpharetta, Ga.), abbreviated
as Neuro Omega.TM. for simplicity. This allows implementing the
proposed open-loop system-identification framework in a clinical
device and testing the system-identification performance in
simulated real-time closed-loop DBS control experiments.
[0077] As an overview, the simulation proceeds as follows. The
Simulated Brain Network 304 generates a simulated ECoG which is
sent to the closed-loop controller 308. The controller 308 includes
the Neuro Omega.TM. 310, which records the simulated ECoG and sends
it to the computer for the controller 312 (PC2). The controller 308
using PC2 312 determines a input pattern (i.e., DBS parameters)
based on the feedback ECoG, and which the Neuro Omega.TM. 310 uses
to generate pulse train DBS waveforms. The Neuro Omega.TM. 310
sends the DBS waveform to the Simulated Brain Network 304, which
generates the simulated ECoG in response to the DBS waveform, thus
closing the loop. A National Instrument Data Acquisition (NI DAQ)
card may be used to implement digital-to-analog conversion (DAC)
and analog-to-digital conversion (ADC). It should be appreciated
that PC1 306 and PC2 312 although logically separate may share
hardware components, and may consist of the same or different
hardware components (e.g., hardware processor, computer memory,
internal bus, etc.). As a concrete example, the closed-loop
simulation testbed 300 was implemented for mood symptom control in
depression, as explained below.
[0078] Simulated Brain Network. FIG. 3B shows a process flow 320
for a simulated brain network 304 receiving DBS pulse train input
from which DBS parameters are calculated or extracted 322 to obtain
u.sub.t. Brain network dynamics under depression may be simulated
by a modified LSSM 324, as follows:
x.sub.t+1=Ax.sub.t+Bu.sub.t+Bd+w.sub.t
y.sub.t=Cx.sub.t+Du.sub.t+Dd+v.sub.t
s.sub.t=Tx.sub.t (5)
[0079] The constant low level of mood characteristic of depression
(i.e., a low mood) may be modeled by a time-invariant unknown
disturbance d. A mood symptom s.sub.t may be quantified by a linear
combination of the hidden network state x.sub.t through matrix T
.di-elect cons.R.sup.1.times.n. This mood symptom in depressed
patients would take a negative value (e.g., s.sub.t=-40, indicating
"depressed") and the closed-loop controller should target a neutral
mood symptom (i.e., s.sub.t=0). The variable d may be set as
d=[T(I-A).sup.-1B].sup..dagger..times.s.sub.depression such that
when u.sub.t=0, the mean value of the steady-state neural mood
symptom is s=s.sub.depression. To simulate a depressed patient, the
value of s.sub.depression may be set s.sub.depression=-40.
Referring to FIG. 3B showing a simulated brain network 302 in a
testbed 300, the modified LSSM is built as an example to simulate
neural activity under depression. The LSSM 324 may be implemented
in a personal computer (PC1) 306 (FIG. 3A). The input u.sub.t may
be simulated as DBS pulse frequency and amplitude, and y.sub.t as
ECoG powers or log-powers. The output y.sub.t is amplitude
modulated 328 according to a dataset of ECoG waveform templates 326
in computer memory. More details are provided under the heading
"Simulation Procedure" below.
[0080] Identification of T and d. To fully identify the depression
LSSM 324, apart from estimating the LSSM parameters n, A, B, C, D,
Q, R and S based on .sup.T={u.sub.t, y.sub.t}.sub.t.sup.T=1 using
methods explained above, an extra open-loop dataset is needed to
estimate T and d. Using the simulation testbed, at rest (i.e., no
stimulation u.sub.t=0) dataset is collected wherein the patient's
mood symptom st is indicated and time-correlated to the ECoG,
resulting in an extra open-loop dataset
T 0 = { s t , y t } .times. T 0 t = 1 . ##EQU00014##
[0081] Based on y.sub.t, the hidden states x.sub.t of the simulated
brain network are estimated by using a Kalman filter (see heading
"Kalman Filter" below). First the mean of the measured
{s.sub.t}.sub.t=1.sup.T.sup.0 is removed and the mean of the
measured {y.sub.t}.sub.t=1.sup.T.sup.0 is removed. Then the neural
state is estimated using the Kalman filter with u.sub.t and d set
at 0. T is estimated from the mean-subtracted observed mood symptom
{s.sub.t}.sub.t=1.sup.T.sup.0 and these Kalman-estimated states
{ x ^ t | t } .times. T 0 t = 1 ##EQU00015##
by least squares regression. The data object s.sub.t=T{circumflex
over (x)}.sub.t|t is defined as the neural signature for mood
symptom. While the mood symptom is assumed to be known for this
open-loop dataset (e.g., by behaviorally measuring it through
self-reported questionnaires), in subsequent closed-loop control,
the controller only observes ECoG and has no direct information of
the true mood symptom s.sub.t. Instead, the controller estimated
the mood symptom by calculating the neural signature s.sub.t based
on the observed ECoG.
[0082] Finally, the controller estimates d from the mean mood
symptom
s _ .apprxeq. 1 T 0 .times. T - 1 t = 0 .times. s t
##EQU00016##
and the mean neural feature
y _ .apprxeq. T - 1 t = 0 .times. y t ##EQU00017##
by solving a set of linear
[0083] equations as described under the heading Identification Of
Unknown Disturbance d below. Note that in practice, identification
time T is typically constrained to be short but not T.sub.0,
because at-rest data can be obtained over a long time period.
[0084] Closed-Loop Controller Design. Once the LSSM 324 is fully
identified, a standard linear-quadratic-Gaussian (LQG) closed-loop
controller 308 is designed to control the network activity. The LQG
controller 308 includes a recursive Kalman state estimator and a
linear-quadratic-regulator (LQR) feedback controller. Derivation of
Kalman filter is straightforward and more detail presented below
under the heading Kalman Filter. More briefly, the LQR controller
design uses an LQR cost function defined as
J .function. ( u t ) = .infin. t = 0 .times. x t T .times. T T
.times. T .times. x + ( u t - u * ) T .times. R .times. w
.function. ( u t - u * ) . ##EQU00018##
Here u* is the steady-state DBS parameters required to maintain the
target level s*=0, which can be computed as u*=d. The optimal
solution of u.sub.t is given by a linear feedback of the state
u.sub.t=-Gx.sub.t+u*. The feedback gain G is given by the Riccati
equation solution. A constraint of positivity of the control
variables may be imposed by bounding this solution. x.sub.t is
given by the Kalman state estimate.
[0085] Computational aspects of the closed-loop controller 308 may
be implemented by a second personal computer 312 (PC2) or by any
suitable hardware processor. Referring to FIG. 3C, operational
aspects of a process 330 for generating DBS parameters u.sub.t is
shown, which may be executed by the computer 312. The computer
calculates ECoG features as feedback to determine the closed-loop
DBS amplitude and frequency (DBS parameters u.sub.t) by the above
LQG feedback control law. Both the LSSM and the closed-loop
controller may run at a common time-step, e.g., 1 s.
[0086] In practical DBS scenarios, recorded neural signals are
contaminated by stimulation artifacts. Typically, stimulation
artifacts have larger amplitudes than neural signals, leading to
low signal to noise ratio. Hence, the hardware processor removes
stimulation artifacts 332 from the recorded ECoG in real time
before calculating neural features y.sub.t 334 from the raw
recorded ECoG. Although FIG. 3C indicates "power features" at block
334, it should be appreciated that any useful signal feature may be
used. In the simulation testbed 300, stimulation artifacts may be
simulated by referencing the generated ECoG to an amplified version
of the input DBS pulse train waveform, such that the actual
recorded signal is a superposition of ECoG and the large
stimulation pulse train artifacts. Closed-loop control requires
real-time removal of the stimulation artifacts, hence the process
330 includes execution of a real-time template-subtraction artifact
removal algorithm 332 before neural feature calculation 334.
[0087] In summary of the foregoing, FIG. 3D illustrates
higher-level aspects of a process 350 for design of a clinical
closed-loop control process. Details may be as explained elsewhere
herein. Open-loop testing is first performed for system
identification at 352. From system identification 352, an LSSM for
the system at hand is estimated 354. Note that in clinical
settings, the LSSM will be estimated for each subject, although
similar LSSMs may be identified for different subjects. In
addition, system identification may need to be repeated for
different closed loop control sessions with the same subject,
because system parameters may drift over time or in response to
variations in measurement, due to stimulation-induced changes or
plasticity, or due to stimulation electrode networks (e.g.,
different equipment in the signal train). After system
identification, the closed-loop controller is designed 356 based on
the estimated system parameters (i.e., the estimated LSSM). At 358,
the closed loop controller is tested experimentally 358 in a
clinical setting, resulting in performance measures for the
experiment. The performance measures indicate effectiveness of the
closed loop controller in mitigating an undesired mood or mental
condition (e.g., depression, anxiety, agitation) or enhancing a
desired mood or mental condition (e.g., confidence, focus) based on
cognitive feedback from the subject assessing the target mental
state during the experiment. In an alternative, the closed-loop
controller may be developed and tested using a simulation procedure
as described below.
[0088] Simulation Procedures. Multiple simulations may be run to
test 350 a proposed 10 identification framework by evaluating both
open-loop identification 352 and closed-loop control 356
performances with subsequent control experiments 358, also
consistent with FIG. 3D. Monte-Carlo simulations may be performed
within computer programs. In addition, or in an alternative,
real-time closed-loop control simulations may be implemented using
the closed-loop simulation testbed 300 with Neuro Omega.TM..
[0089] Monte-Carlo Simulations. The proposed IO identification
framework may be tested using numerical simulations by a custom or
commercial analysis computer program, e.g., Matlab.TM.. An
advantage of Monte-Carlo experiments is that the experiments can be
completed in a short amount of time. The IO identification
framework may be tested with BN input pattern 402 (FIG. 4A), where
each DBS parameter is an independent BN sequence that satisfies
equation (2), for example with T.sub.sw=1 s, DBS amplitude levels
L.sub.1=2 mA and L.sub.2=6 mA, and DBS frequency levels L1=50 Hz
and L2=100 Hz. The upper solid line 401 represents the power
spectral density of the BN input pattern, which is white across the
frequency band. Three alternative input patterns are compared with
the BN input pattern:
[0090] (i) rectangular pattern; each parameter has a rectangular
shape:
u t = L , .A-inverted. t .di-elect cons. [ 1 4 .times. T , 5 4
.times. T ] .times. .times. .times. and .times. .times. u t = 0 ,
.A-inverted. t .di-elect cons. [ 0 , 1 4 .times. T ] [ 5 4 .times.
T , 3 2 .times. T ] , ##EQU00019##
where
L = L 1 + L 2 2 . ##EQU00020##
Here the effective stimulation time is still T, but we set
aside
T 4 ##EQU00021##
before stimulation and after stimulation to include the transition
periods. [0091] (ii) single-sinusoid pattern; each parameter has a
sinusoid shape with a single frequency f chosen at random:
u.sub.t=L+L sin(2.pi.f t). [0092] (iii) multiple-sinusoid pattern;
each input parameter is a superposition of six sinusoids:
[0092] u t = L + L 5 .times. 6 k = 1 .times. sin .function. ( 2
.times. .pi. .times. f k .times. t ) , ##EQU00022##
with the frequencies f.sub.k's chosen at random.
[0093] For a sample Simulated Brain Network, random LSSM parameters
may be generated as follows. First determine the model dimensions.
For example, model order may be randomly selected n from {2, 3, . .
. , 20}, the number of stimulation input channels m.sub.1 from {1,
2, . . . , 10}, and the number of recorded output ECoG channels
from {m.sub.1,m.sub.1+1, . . . , 3m.sub.1} since in practice there
are more output channels than stimulation channels. Then, the A, B,
C, and D matrices may be randomly generated, for example, by the
Matlab.TM. built-in function drss such that A is stable
(eigenvalues inside the unit disk). Next, the noise covariance
matrix is defined as a random positive-semidefinite matrix
.epsilon..times.U.sup.TU, where U is a full-rank (n+k).times.(n+k)
random Gaussian matrix with unit variance, and .epsilon. is
randomly selected from [0.001, 0.05]. Finally, T may be set as
random Gaussian matrix, and d may be set as
d=[T(I-A)-.sup.1B].sup..dagger..times.s.sub.depression, with
s.sub.depression=-40.
[0094] In open-loop identification experiments, an LSSM may be
identified for each of the input patterns, wherein T=120 mins in
.sup.T and T.sub.0=10 mins in .sup.T.sup.0. The identified LSSMs
may then be used design closed-loop controllers, referred to herein
as a BN controller, rect controller, single-sin controller and
multi-sin controller. A target objective is defined, for example,
for the depression example the closed-loop controller may target a
neutral mood s*=0. Lower- and upper-bound limits on the DBS
frequency may be defined, for example 0 to 150 Hz, and for the DBS
current amplitude 0 to 8 mA.
[0095] As the comparison with "worst case" baseline in
system-identification, stable LSSMs can be generated using the same
method as in the Simulated Brain Network. These LSSMs may be
referenced herein as "rnd" (random), and used to estimate the
chance level system-identification error. As the comparison with
"best case" baseline in closed-loop control, an oracle controller
derived from the true LSSM used in Simulated Brain Network may be
used. The oracle controller has the best possible closed-loop
control performance.
[0096] Both system-identification and closed-loop control can be
tested by running Monte-Carlo experiments across a large number
(e.g., 100) of Simulated Brain Networks. In addition, for each of
the Simulated Brain Networks, multiple system-identification and
closed-loop control trials (e.g., ten each) may be performed with
independent noises. Different controllers may be tested for
comparative purposes. For example, for a particular Simulated Brain
Network, Wilcoxon rank-sum tests may be run across the multiple
closed-loop control trials. The performance of these multiple
trials may be averaged as the final control performance of a given
controller for this Simulated Brain Network. Wilcoxon signed-rank
tests may similarly be performed to compare performance across the
100 Simulated Brain Networks.
[0097] Closed-Loop Simulations With Neuro Omega.TM.. In addition,
real-time closed-loop simulations can be performed using the
simulation testbed with Neuro Omega.TM.. For example: [0098] (i) To
show the performance of BN-modulated DBS waveform when
identification time is not constrained, a long identification time
may be selected, i.e., T=120 mins. First, an ideal case where there
exist no stimulation artifacts is simulated. The simulated
controller in this case may be referred to herein as the "BN"
controller. [0099] (ii) Next stimulation artifacts and implement
artifact removal algorithm may be added in both open-loop
identification and closed-loop control. The simulated controller
with this modification may be referred to herein as the
"BN-with-artifact" controller. [0100] (iii) To compare the
performance of BN-modulated DBS waveform and GBN-modulated DBS
waveform when identification time is restricted, two more input
patterns may be tested. The first one is a shorter BN input pattern
with T=20 mins. The resulting controller may be referred to herein
as the "BN-20 mins" controller. The second input pattern may be a
GBN input pattern with T=20 mins and
[0100] p = 1 - T s .times. w T ^ c . ##EQU00023##
The resulting controller may be referred to herein as the "GBN-20
mins" controller. In this comparison, we also simulate the
stimulation artifacts.
[0101] T.sub.sw may be set to 1 s in both BN input pattern and GBN
input pattern, and the parameter levels the same as described above
for the Monte Carlo simulation. For simplicity, for the Simulated
Brain Network, a single stimulation channel may be simulated with a
choice of the effective DBS parameters u.sub.t as the pulse
frequency and amplitude. Therefore, the LSSM input dimension m=2=1
stimulation channel.times.2 DBS parameters. The output neural
features y.sub.t may be selected as the ECoG powers or log-powers
of plural frequency bands, for example, 1-4 Hz, 5-8 Hz, 9-16 Hz,
17-32 Hz, and 33-64 Hz. Two channels of ECoG may be simulated.
Hence, in this example the LSSM output dimension k=10=2 ECoG
channels.times.5 power features.
[0102] In total, two Simulated Brain Networks were tested by the
authors hereof with independently generated random LSSMs. The first
Simulated Brain Network (Simulated Brain Network 1) had a random
underlying LSSM of order 3, with poles at 0.97, 0.95 and 0.93, and
noise level .epsilon.=0.005. The second Simulated Brain Network
(Simulated Brain Network 2) had a random underlying LSSM of order
of order 4, with poles at 0.95, 0.90, 0.85 and 0.75, and a higher
noise level .epsilon.=0.02. The closed-loop control design settings
were the same as the previous section. For each of the Simulated
Brain Network and each of the above four controllers, i.e., BN
controller, BN-with-artifact controller, BN-20 mins controller and
GBN-20 mins controller, the authors ran 10 trials of closed-loop
control experiments and conducted Wilcoxon rank-sum tests to
compare performance across the 10 trials.
[0103] Performance Measures. The brain network IO identification
framework may be evaluated both in open-loop identification
experiments and in closed-loop control experiments. In open-loop
identification, the relative estimation error of the IO dynamics
can be quantified as
E I .times. O .times. - ^ 2 2 , ( 6 ) ##EQU00024##
where denotes the true LSSM and denotes the estimated LSSM. The
norm .parallel..parallel..sub.2 is the .sub.2 norm of a dynamical
linear system (see heading Definition of System Norm below).
Similarly, the relative estimation error of the noise covariance
can be quantified as
E n .times. o .times. i .times. s .times. e .times. - 2 2 , ( 7 )
##EQU00025##
where represents a dynamical system describing the IO dynamics from
the noise terms w.sub.t and v.sub.t to the output y.sub.t in (1).
is termed noise dynamics, and is closely related to the noise
covariance (Definition of System Norm). denotes the estimated noise
dynamics.
[0104] In closed-loop control experiments, various metrics may be
used. Steady-state performance may be evaluated by computing error
et between a controlled underlying mood symptom s.sub.t and a
target level s*=0, i.e., e.sub.t=s.sub.t. The steady-state bias
BIAS may be defined as
BIAS=mean(e.sub.t), (8)
and the steady-state standard deviation STD as
STD=std(e.sub.t), (9)
The mean and standard derivation may be calculated across data
points at steady state. Transition time TranT may be defined as the
time it takes to complete 95% of the transition from the steady
state before closed-loop control to the steady state with
closed-loop control. Estimation performance error EstE of the
estimated neural signature s.sub.t, with respect to the true mood
symptom s.sub.t, may be calculated as
EstE = t = 0 N - 1 .times. ( s t - s ^ t ) 2 t = 0 N - 1 .times. s
t 2 ( 10 ) ##EQU00026##
where s.sub.t=T{circumflex over (x)}.sub.t|t is the estimated
neural signature with {circumflex over (x)}.sub.t|t given by the
Kalman filter and N is the total closed-loop control experiment
time.
[0105] Results: BN Input Pattern Achieved Accurate Open-Loop Model
Identification: Monte Carlo Simulations. Monte Carlo
system-identification simulations were performed to compare the
four input patterns described in connection with FIG. 4A above.
Power spectrums of each input pattern were estimated by Welch's
method, and show that the BN input pattern has approximately white
spectrum. The spectrums of other input patterns are concentrated
around particular frequencies. To see an example of how the input
patterns affected identification results, a sample Simulated Brain
Network was generated and four LSSMs using the four input patterns
were identified. The estimated IO transfer functions from the 1st
input channel to the 1st output channel for a BN input pattern is
shown as the Bode plot 410 in FIG. 4B. Similarly, the Bode plot 420
compares transfer functions for a rectangular input pattern, Bode
plot 430 for a single-sine input pattern, and Bode plot 440 for a
multi-sine input pattern. Compared with other input patterns, the
BN input pattern led to a more accurate estimate of the underlying
IO dynamics with less bias and variance.
[0106] The above observations in Monte-Carlo simulations were
quantified. IO estimation error of the BN input pattern was much
smaller than other input patterns as shown in the boxplots 450 of
FIG. 4C (median E.sub.IO, BN: 0:0094; rect: 0:9210; single-sin
1:0831; multisin 0:9451; rnd 1:4114). In FIG. 4C, the lower and
upper bound of the boxes represent the 25th and 75th percentiles of
the absolute error distribution and the middle line in each box
represents the median. Whiskers represent the 95th and 5th
percentile of the absolute error distribution. The same holds for
all subsequent box plots. In estimating the IO dynamics, BN is
advantageous because it has white spectrum and can excite the
entire frequency band with equal energy and without bias towards
specific bands. The BN input pattern also estimated the noise
dynamics successfully, with much smaller error than the chance
level performance as shown in the boxplots 460 of FIG. 4D (median
E.sub.noise, BN: 0:1692 v.s. rnd: 1:0742). The four input patterns
had comparable estimation errors of the noise dynamics, because
noise dynamics were driven by w.sub.t and v.sub.t, which were
independent of the input patterns u.sub.t.
[0107] Results: BN controller achieved precise closed-loop control
of mood symptom: Monte Carlo simulations. The brain 10 dynamic
identification framework was further validated by evaluating how
the identified models influenced closed-loop DBS control. FIGS.
5A-D show results of a control session for a sample Simulated Brain
Network. The BN controller 500 controlled the mood symptom s.sub.t
to follow the target closely, as shown in the symptom target graph
502 and DBS frequency graph 504 in FIG. 5A (BIAS=-0.05 and
STD=0.98) by actively changing the DBS frequency and amplitude
according to neural feedback. The estimated neural signature
s.sub.t closely followed the mood symptom as shown in FIG. 5A
(EstE=0.03). In addition, the BN controller performed similar to
the best-case oracle controller as also shown at FIG. 5A, 502.
Closed-loop control performance of the BN controller (i.e., BIAS
540, STD 542, TranT 544 and EstE 546, see above) as shown in FIG.
5E was not significantly different from the oracle controller
across the 10 closed-loop control trials with independent noises
(BN v.s. oracle; Wilcoxon rank-sum tests, P>0.15 for all
performance measures).
[0108] In contrast, the other three controllers had a large bias in
controlling the mood symptom s.sub.t at steady state. Charts 512,
514 of FIG. 5B show a large bias for a rectangular input controller
510, compared to target and to oracle. Charts 522, 524 of FIG. 5C
show a similar bias for a single sine controller 520. For the
multi-sine controller 530, a similar bias appears in charts 532,
534 of FIG. 5D. Also, FIGS. 5B-D show the estimated neural
signatures s.sub.t were biased from the underlying mood symptom st.
Compared with the oracle controller, the performance of these three
controllers was significantly worse, as shown in FIG. 5E (rect v.s.
oracle, single-sin v.s. oracle, multi-sin v.s. oracle: P<0.05
for all performance measures).
[0109] We further conducted Monte Carlo simulations in which we
controlled 100 independent Simulated Brain Networks, the results of
which are depicted in FIG. 5F, wherein chart 550 show BIAS, chart
552 shows STD, chart 554 shows TranT, and chart 556 shows EstE. The
performance of the BN controller and the oracle controller was not
significantly different across the 100 Simulated Brain Networks for
BN v.s. oracle (Wilcoxon signed-rank tests, P>0.35 for all
performance measures). For the other three controllers, the
performance was significantly worse than the oracle controller
(Wilcoxon signed-rank tests, P<0.05 for all performance
measures).
[0110] Combined, the results demonstrate that i) input pattern in
open-loop system-identification has critical influence on
subsequent closed-loop control performance, and ii) that the novel
BN input pattern can achieve accurate IO dynamics identification
and thus, enable accurate closed-loop control of the mood
symptom.
[0111] Results: BN Controller Achieved Accurate Closed-Loop Control
Of Mood Symptom: Neuro Omega.TM. Simulations. Simulations using the
closed-loop simulation testbed with Neuro Omega.TM. were also
conducted. For both Simulated Brain Network 1 and Simulated Brain
Network 2, the controlled mood symptom using BN controller followed
the target tightly at steady-state and was close to the oracle
controller as shown in chart 500, FIG. 6A (Brain Network 1) and
chart 620, FIG. 6B (Brain Network 2). The performance of BN
controller and oracle controller was not significantly different
for both brain networks across the 10 closed-loop simulation trials
as shown in the boxplots 610, 612, 614, 616 of FIG. 6B (Simulated
Brain Network 1, BN v.s. oracle, Wilcoxon rank-sum tests, P>0.4
for all performance measures; and boxplots 630, 632, 634, 636 of
FIG. 6D (Simulated Brain Network 2, BN v.s. oracle; Wilcoxon
rank-sum tests, P>0.5 for all performance measures).
[0112] Results: BN Controller Achieved Accurate Performance In Face
Of Stimulation Artifacts: Neuro Omega.TM. Simulations. Simulation
of a more practical case where there existed stimulation artifacts
in the simulation testbed, as described herein above, showed
continued accurate performance of the BN Controller. By
implementing real-time artifact removal in the controller (PC2,
FIG. 3C) for both Simulated Brain Networks 1 and 2, the controlled
mood symptom followed the target closely as shown in FIGS. 6A, 6C,
and were again close to the mood symptom controlled by the oracle
controller. Across the 10 closed-loop simulation trials, there
existed no significant differences between the BN-with-artifact
controller and the oracle controller for both Simulated Brain
Networks as shown in FIG. 6B (Simulated Brain Network 1,
BN-with-artifact v.s. oracle, Wilcoxon rank-sum tests, P>0.2 for
all performance measures) and in FIG. 6D (Simulated Brain Network
2, BN-with-artifact v.s. oracle; Wilcoxon rank-sum tests, P>0.15
for all performance measures).
[0113] Results: GBN-Modulated DBS Waveform Reduced
System-Identification Time: Neuro Omega.TM. Simulations. Results
above demonstrate that BN-modulated DBS waveform enables accurate
system-identification and closed-loop control when the
identification period is long (T=120 mins). However, there may be
constraints on how long the brain can be stimulated in open-loop
for system-identification. Tests of how GBN-modulated DBS waveform
may improve BN-modulated DBS waveform for time-constrained
scenarios where T was restricted to be short (20 mins) were
conducted. Simulations in the simulation testbed with Neuro
Omega.TM. were used to compare the BN-20 mins controller and the
GBN-20 mins controller, where stimulation artifacts were added in
both open-loop system-identification and closed-loop DBS
control.
[0114] The BN-20 mins controller was tested in Simulated Brain
Network 1. Its steady-state bias, transition time and neural
signature estimation error were significantly worse than the oracle
controller across the 10 closed-loop control trials as shown in
boxplots 730, 732, 734 and 736 of FIG. 7B (BN-20 min v.s. oracle;
Wilcoxon rank-sum tests: BIAS P=1.8.times.10.sup.-4, TranT P=0.006
and EstE P=1.8.times.10.sup.-4). The GBN-20 mins controller using
an input pattern for different time constants as shown in charts
700, 720 FIGS. 7A, 7C. The time constant of the underlying network
dynamics was estimated using a 2 mins on-off stimulation (see
disclosure above, note this extra time was much shorter compared
with the previous 120 min BN-modulated DBS waveform). The estimated
network time constant ({circumflex over (T)}.sub.c) was 25 s. In
comparison, the true time constants corresponding to the four poles
of the LSSM used in Simulated Brain Network 1 were 32.8 s, 19.5 s,
and 13.7 s (T.sub.c=-1/ln(p.sub.R), where p.sub.R is the real part
of a pole), with an average of 23 s. Hence the time constant
estimate was close to the true average time-constant. Based on this
estimate, we designed a GBN with p=1-1/25=0.96, consistent with
Equation 4 above; GBN pattern shown in FIG. 7A). Compared with
oracle controller, GBN-20 mins controller showed no significant
difference across the 10 closed-loop control trials as shown in
FIG. 7B (GBN-20 mins v.s. oracle; Wilcoxon rank-sum tests:
P>0.25 for all performance measures).
[0115] The reason for suboptimal performance of the BN20 mins
controller is that the underlying brain network was dominated by
slow dynamics. Hence BN was changing too frequently compared with
the inherent dynamics. This led to poor excitation and estimation
of low frequency components of the dynamics. Inaccurate estimation
of d resulted in large steady-state bias; inaccurate estimation of
T resulted in a poor estimation of the neural signature and further
influenced the transition time via the control cost function.
Compared with the BN input pattern with p=0.5 (FIG. 2A), the GBN
input pattern had p=0.96 and thus made changes much less frequently
(FIG. 7A), resulting in more low frequency components. This led
system-identification to weigh the low frequency bands more. These
bands largely coincided with the important dynamics of the
underlying brain network, leading to more accurate model estimation
than was possible in the BN-20 mins case.
[0116] Simulated Brain Network 2 was similarly tested, wherein the
BN-20 mins controller also had degenerated performance and
performed significantly worse than the oracle controller across the
10 closed-loop control trials, as shown in boxplots 730, 732, 734,
736 of FIG. 7D (BN-20 min v.s. oracle; Wilcoxon rank-sum tests:
oracle v.s. BN-20 mins: BIAS P=0:0017, TranT P=0:0150 and EstE
P=0:0073). An estimated network time constant of 10 s was obtained
from a 3 mins on-off stimulation, which was close to the average
time constants of the four poles (9.68 s). Therefore, for Simulated
Brain Network 2, GBN took parameter p=1-1=10=0:90 (Eq. 4). The GBN
pattern is shown in chart 720, FIG. 7C. Consistent with the results
for Simulated Brain Network 1, the closed-loop performance of the
GBN-20 mins controller was similar to the oracle controller across
the 10 closed-loop control trials as shown in FIG. 7D (GBN-20 mins
v.s. oracle; Wilcoxon rank-sum tests: P>0:3 for all performance
measures).
[0117] Results: BN Controller Was Robust To Unknown Mood
Disturbance Levels: Neuro Omega.TM. Simulations. The robustness of
the proposed system-identification framework to unknown levels of
inherent depressed mood was examined. Six closed-loop experiments
were ran with the Neuro Omega.TM. simulation testbed under various
mood disturbance levels, stimulation artifacts were also simulated.
In system-identification, a BN-modulated DBS waveform with T=120
mins was implemented.
[0118] For both Simulated Brain Network 1 and Simulated Brain
Network 2, we simulated three scenarios, with s.sub.depression=-20,
-30, -40, representing a progression from less depressed patients
to more depressed patients. Each of these closed-loop control
sessions lasted for 800 s, where the first 210 s were stimulation
off, followed by 590 s where closed-loop stimulation was on. The BN
controller achieved transitions without overshoot and precise
control of the mood symptom at steady-state across these 6
simulations. A first simulation is shown in charts 800 of FIG. 8A,
with chart 802 showing mood symptom control and chart 804 showing
DBS frequency overtime (1.sup.st Brain Network, disturbance level
-20). FIG. 8B similarly shows in charts 810, 812, 814 a second
simulation for the 1.sup.st Brain Network, disturbance level -30.
FIG. 8C shows in charts 820, 822, 824 a third simulation for the
1.sup.st Brain Network, disturbance level -40. A fourth simulation
for the Simulated Brain Network 2, disturbance level -20, is shown
charts 830 of FIG. 8D with chart 832 showing mood symptom control
and chart 834 showing DBS frequency over time. FIG. 8E similarly
shows in charts 840, 842, 844 a fifth simulation for the 2nd Brain
Network, disturbance level -30. FIG. 8F shows in charts 850, 852,
854 a sixth simulation for the 2nd Brain Network, disturbance level
-40. In addition, the estimated neural signature s.sub.t followed
closely the true mood symptom s.sub.t (see FIGS. 8A-F). By
comparing the charts in FIGS. 8A-F across columns, we see that the
BN controller was robust to different levels of unknown mood
disturbances. By comparing the subfigures in FIGS. 8A-F across rows
we see that the BN controller was robust to different underlying
brain networks.
[0119] Discussion/Summary. A new computational framework has been
presented herein to model and identify IO dynamics of brain
networks in response to electrical stimulation. Implementation of a
real-time clinical closed-loop simulation testbed using Neuro
Omega.TM. is used to validate the framework. In the computational
framework, 1) general LSSMs are built to describe the IO dynamics,
identify a low-dimensional brain state, and enable closed-loop
control, 2) a new open-loop input DBS/stimulation waveform-a pulse
train modulated by BN parameters, or BN input patterns that is both
critical for model identification and consistent with current DBS
safety conventions is developed and tested, and 3) the BN-modulated
DBS waveform is generalized to a GBN-modulated DBS waveform to
reduce the time needed for open-loop brain stimulation to identify
the models. In the closed-loop testbed, apparatus and methods are
described for (i) open-loop system identification using the
(G)BN-modulated DBS waveform, and (ii) closed-loop control of the
brain network. The present disclosure uses depression as an
example, but other brain states may similarly be controlled. Within
this testbed, stochastic noises, unknown disturbances, and
stimulation artifacts to simulate were incorporated as closely as
possible to a real-world scenario and resulting robustness was
assessed.
[0120] The modeling framework was validated by two sets of
experiments: (i) open-loop system-identification experiments to
assess closeness of the estimated model in the Neuro Omega.TM.
testbed to the true brain network model, and (ii) closed-loop
control experiments to evaluate mood symptom control in depression
within this testbed. It was shown that 1) the novel design of the
BN input pattern is critical for accurate identification of brain
networks; the proposed BN input pattern achieved accurate model
parameter identification, outperforming other suboptimal input
patterns; 2) the novel design of the BN input pattern is critical
for accurate closed-loop control of brain networks; in contrast to
other input patterns, the closed-loop controller built from BN
identification achieved tight control of the mood symptom, with a
steady-state and transient performance similar to the best possible
performance (oracle controller); 3) to enable short open-loop brain
stimulation for model identification, the new GBN input pattern can
improve the performance of the BN input pattern by incorporating
the time-constant information of the brain network; this
time-constant information can in turn be estimated in a short
on-off stimulation session.
[0121] LSSM As A Model Structure For Brain Network IO
Identification And Closed-Loop Control. In the present disclosure,
LSSMs are used to describe the IO dynamics of brain networks to
enable the design of closed-loop DBS systems. While neuronal
dynamics are in general nonlinear, there are multiple reasons to
select an LSSM for the purpose of closed-loop DBS.
[0122] First, closed-loop DBS requires the use of IO models that
are amenable to efficient identification, and for which performance
guarantees can be provided for safety. This is especially the case
given the high-dimensional neural recordings (e.g., hundreds of
ECoG recording electrodes) that need to be modeled and used as
feedback in many DBS applications. LSSM provides a model structure
that satisfies these conditions. Design of a novel (G)BN-modulated
DBS waveform for brain network IO identification was accomplished
by combining the requirements for DBS waveforms and the theoretical
requirements that exist for LSSM model identification (i.e., white
spectrum inputs). This is in contrast to nonlinear models for which
there exist no such global guarantees, hindering an optimal DBS
input design. In addition, to estimate LSSMs, projection and
subspace identification algorithms can be exploited, which are
numerically more stable and have lower computational burdens than
recursive identification algorithms for nonlinear systems, e.g.,
prediction-error-methods. The high-dimensional and large amount of
identification data can cause both stability and computation burden
problems for nonlinear system identification.
[0123] Second, a brain network IO model for DBS should be amenable
to closed-loop control. LSSM is a powerful model structure that
allows optimal closed-loop controller design. Many well-established
estimation (e.g., Kalman estimator, robust H.infin. estimator) and
control (e.g. linear-quadratic-regulator, model predictive control)
algorithms can be used for LSSMs; these designs have led to
successful closed-loop controllers for numerous industrial
processes despite their potential inherent non-linearity.
Computational neural network models consisting of nonlinear neural
oscillators have been used in simulations to study the effects of
open-loop high frequency DBS, or to explore possible closed-loop
DBS strategies. However, these models are not suitable for robust
and accurate closed-loop controller design. For example,
simulations show that based on nonlinear IO models, the choice of
feedback gain that guarantees closed-loop stability could only be
obtained through trial and error simulations since no closed-form
solutions existed. In contrast, LSSM modeling can lead to an
efficient and practical approach to model the dynamics of
electrophysiological time-series (e.g., spiking rates, LFP and
ECoG) and the effect of DBS on these dynamics; this in turn enables
the development of closed-loop controlled DBS systems.
[0124] Third, linear models have been used to describe spontaneous
(i.e., without external input) neural dynamics at the large
macroscopic scale of neural populations, e.g.,
electroencephalography (EEG) or magnetoencephalography (MEG). More
specifically, even though cell level neuronal dynamics are highly
nonlinear, macroscopic level spontaneous neural activity emerging
from complex collections of numerous nonlinear neuronal dynamic
processes can be described by linear models for different purposes.
Linear autoregressive models have been built to study EEG dynamics,
and linear models used to characterize MEG dynamics for estimation
of brain source current. Linear models have also been used to
quantify rotational spontaneous neural dynamics during reaching
behaviors. While a goal includes modeling the input-output neural
dynamics in response to DBS rather than spontaneous neural
dynamics, these prior studies provide evidence that linear models
could also be sufficient for IO identification.
[0125] Finally, there are two flexible design choices that can
further mitigate the effect of potential nonlinearities on the
present modeling framework. First, we have the flexibility to pick
the model order. It is well-known that picking a linear model of
higher order can sufficiently capture nonlinear dynamics in many
cases. Second, we can use a time-varying piece-wise linear model
instead of a time-invariant linear model. In this case, we would
divide the DBS parameter space into several regimes, build a
different linear model for each regime, identify these models using
BN input patterns with appropriate DBS parameters within these
regimes, and finally derive piece-wise linear models and feedback
controllers. Finally, it is well-known that feedback control
linearizes a nonlinear system, further making the case for the use
of LSSM for the purpose of closed-loop control.
[0126] BN-Modulated DBS Waveform For Identification Of Brain
Network IO Dynamics. System-identification is a crucial component
in designing model-based controllers that interact with the brain.
For example, as shown in closed-loop BMIs for anesthesia control,
or in closed-loop motor BMIs. Data-driven open-loop system
identification requires informative IO datasets. This critically
requires an optimal input signal design. In traditional linear
industrial and chemical process identification, BN inputs have been
used since they have white spectrum and are convenient to operate.
BN behavioral input stimulus have also been used to study
biological systems such as the auditory nervous system, the
horizontal vestibuloocular reflex system and the motor system.
[0127] To enable identification of brain network IO dynamics in
response to DBS, and consequently to develop closed-loop DBS
systems, we designed a novel open-loop DBS waveform by utilizing
the concept of BN. The DBS waveform needs to be charge-balanced in
a short time and hence typically must consist of biphasic pulses.
Moreover, it is the DBS parameters such as pulse amplitude and
frequency that affect neural and clinical outcomes. Hence a direct
BN DBS waveform is not useful for brain stimulation. The main
innovation in designing BN-modulated DBS waveform is that the
neural dynamics be related to DBS parameters instead of the DBS
waveform itself by designing a pulse train whose DBS parameters are
stochastically changed according to a BN sequence. This waveform
can incorporate as its building block any conventional DBS pulse
shape such as monophasic or biphasic pulses, or even specialized
DBS pulse shapes that can reduce stimulation artifact duration.
This flexibility allows the BN-modulated DBS waveform to generalize
across applications given that different pulse patterns can lead to
different DBS effects.
[0128] BN-modulated DBS waveform may require a long identification
time in some scenarios. However, the duration of open-loop brain
stimulation to identify models should be restricted for practical
and safety reasons. The BN-modulated DBS waveform is thus extended
to a GBN-modulated DBS waveform by estimating the time-constant of
the underlying brain network, and using this estimate in the
selection of the switching probability of the waveform parameters.
This new DBS waveform allows for a significant reduction in
identification time without a cost on accuracy.
[0129] Towards Clinical Closed-Loop DBS Experiments. The proposed
LSSM modeling and BN identification methods have been verified by
developing a real-time closed-loop BMI simulation testbed using the
clinical Neuro Omega.TM. microelectrode recording and stimulation
system. This new simulation testbed allows for extensive
simulations that are essential to ensure reliability and safety
before moving towards clinical human experiments. Numerical
simulations have also been used to test several closed-loop DBS
designs. However, these prior works only carried out computer
program simulations and did not implement a testbed using an actual
clinical neurophysiological recording and stimulation system. To
enable realistic simulations, we built this testbed using the same
clinical system that would be used in human experiments, and
incorporated in our simulations stochastic noises, disturbances,
time-delays, and recording stimulation artifacts that are difficult
to mimic purely using computer program. This closed-loop DBS system
fully ensembles a real clinical scenario except that the ECoG
signal was generated by a simulated brain network rather than a
human brain. While the testbed was used for validation of the LSSM
modeling and BN identification frameworks, the testbed itself is
general and can be used to simulate closed-loop treatment of other
neural disorders by changing the simulated brain model accordingly,
e.g., using the nonlinear models used for PD or epilepsy.
[0130] By validating the performance of BN identification and the
corresponding closed-loop controller within the clinical simulation
testbed, this work paves the way for active efforts to develop
effective closed-loop treatments for neurological disorders such as
depression and anxiety and implementing them in human clinical
experiments. Two important directions towards clinical testing may
be pursued:
[0131] (i) Identification of network neural signatures for various
neurological disorders. Neural signatures relate neural dynamics to
patient behavioral symptoms, and are necessary feedback signals for
closed-loop control. There has been extensive work in studying
abnormal neural oscillations or couplings for epilepsy, Parkinson's
disease and dystonia. However, the mechanisms underlying these
abnormal neural activities and the choice of a neural signature are
still under extensive investigation. Moreover, neural signatures
for other neurological disorders such as neuropsychiatric disorders
including depression and anxiety are largely unknown. The LSSM
modeling framework proposed here can characterize neural dynamics
at the network level. Hence it is important to investigate whether
LSSM modeling can identify neural signatures for neurological
disorders such as mood disorders including depression and
anxiety.
[0132] (ii) BN identification and model-based closed-loop control
in animals and humans. Having established the performance of the
proposed methods in a clinical simulation testbed, a next step is
to apply BN stimulation in animal and human subjects using the
exact protocol used in the Neuro Omega.TM. simulations presented
herein. The new BN IO identification framework allows investigation
of the effect of electrical stimulation on neural activity and the
neural signatures of disease. The identified IO model will both
facilitate deeper understanding of the dynamics of brain networks,
and better design of closed-loop controllers.
[0133] In an aspect, the use of BN/GBN waveforms as described
herein for brain stimulation can also be applied to non-electrical
stimulation inputs such as for example optogenetic stimulation. For
optogenetics the parameters to switch over are for example the rate
and/or width and/or amplitude of the light pulses. Other aspects of
BN/GBN identification and use of the identified waveform for
closed-loop control may otherwise be as described herein for
electrical stimulation, when applied to optogenetic
stimulation.
[0134] Kalman Filter And LQR Controller Design. The Kalman
estimator gives optimal estimate of x.sub.t via the following
recursion
{ x ^ ( t + 1 | t ) = A .times. x ^ t | t + B .times. u t + B
.times. d _ x ^ ( t + 1 | t + 1 ) = x ^ t | t + L ( y t + 1 - C
.times. x ^ t t - D .times. u t - D .times. d _ ( 11 )
##EQU00027##
where the gain matrix L is the solution of the algebraic Riccati
equation [53] (given that the pair [A,C] is observable)
L=PC.sup.TR.sup.-1-PC.sup.T (CPC.sup.T+R)CPC.sup.TR.sup.-1, where
the matrix P satisfies the Ricatti equation
P=APA.sup.T+Q-(APC.sup.T+S)(CPC.sup.T+R).sup.-1(CPA.sup.T+S). Here,
we set the initial estimate as {circumflex over (x)}(-1|-1)=0.
[0135] The LQR controller takes as feedback the Kalman state
estimate and decides on the simulation parameters u.sub.t to take
the diseased brain state to a healthy level, which we quantify in
terms of the neural signature as s*=0. LQR algorithm minimizes the
quadratic cost function
J(u.sub.t)=.SIGMA..sub.t=0.sup..infin.x.sub.t.sup.TT.sup.TTx.sub.t+(u.su-
b.t-u*).sup.TR.sub.w(u.sub.t-u*) (12)
where u* is the steady-state stimulation parameters required to
maintain the target level s*=0, and R.sub.W is a positive definite
matrix chosen depending on the desired system response. The optimal
solution of u.sub.t is given by a linear feedback of the state
u.sub.t=-Gx.sub.t+u* (13)
The feedback gain G is given by the Riccati equation solution
(given that the pair [A,B] is controllable),
G=(R.sub.W+B.sup.THB).sup.-1B.sup.THA, where H satisfies the
Riccati equation
H=A.sup.THA+T.sup.TT-A.sup.THB(R.sub.W+B.sup.THB).sup.-1B.sup.THA.
Also, it's easy to verify that to keep s*=0, the sufficient
condition is u*=d. In addition, we impose the constraint of
positivity of the control variables by bounding the solution in
(13). In online implementation, the value of x.sub.t is given by
the state estimate {circumflex over (x)}.sub.t|t in (11). Also, we
note that in all simulations, we select the LQR weighting matrix
as
R W = [ 1 .times. 0 - 3 0 0 1 .times. 0 - 2 ] . ##EQU00028##
[0136] Identification Of Unknown Disturbance d. We estimate d based
on
T = { s t , y t } .times. T - 1 t = 0 . ##EQU00029##
At steady state, the contribution of the constant disturbance is on
the mean of s.sub.t and y.sub.t. Hence, we take the mean of both
sides in (5) and get
x=Ax+B +Bd
y=Cx+D +Dd
s=Tx
In this data set, we have u.sub.t=0. Plugging this into the above
equations, and by rearranging the variable, we have the following
set of linear equations with known variables s and y.sub.t and
unknown variables x and d:
[ A - I B C D T 0 ] .function. [ x _ d _ ] = [ 0 y _ s _ ]
##EQU00030##
Then, we estimate the constant disturbance as the minimum-norm or
least-squares solution of the above equations.
[0137] Definition Of System Norm .sub.2. We first derive an
equivalent form of the LSSM (1). Assuming the LSSM (1) is
observable, an equivalent description of the input-output (IO)
relationship and second order noise statistics is
X.sub.t+1=Ax.sub.t+Bu.sub.t+Ke.sub.t
y.sub.t=Cx.sub.t+Du.sub.t+e.sub.t (14)
with K=(APC.sup.T+S)(CPC.sup.T+R).sup.-1, where matrix P satisfies
the Ricatti equation
P=APA.sup.T+Q-(APC.sup.T+S)(CPC.sup.T+R).sup.-1(CPA.sup.T+S). In
addition, E[e.sub.ie.sub.j.sup.T]=VS.sub.ij where V=CPC.sup.T+R.
The above LSSM captures two important dynamics in neural activity.
The first one is IO dynamics and the second one is noise dynamics.
We further describe these two relationships more compactly as
follows. By applying transform (z-transform) to (14), we have
zX(z)=Ax(z)+BU(z)+KN(z)
Y(z)=Cx(z)+DU(z)+N(z) (15)
where X(z)=(x.sub.t), U(z)=(u.sub.t), Y(z)=(y.sub.t), and
N(z)=(e.sub.t). By solving the first equation for X(z), and
plugging the solution into the second equation, we have
Y(z)=G(z)U(z)+H(z)N(z), (16)
where
G(z)=C(zI-A).sup.-1B+D, (17)
and
H .function. ( z ) = C .function. ( z .times. I - A ) - 1 .times. K
.times. V 1 2 + V 1 2 ( 18 ) ##EQU00031##
Here, we define G(z) as the IO dynamics and H(z) as the noise
dynamics. The system norm is then defined as
.parallel..parallel..sub.2=
1 2 .times. .pi. .times. .intg. .infin. - .infin. .times. Tr ( G
.function. ( e j .times. w ) H .times. G .function. ( e j .times. w
) .times. d .times. .omega. ##EQU00032##
where Tr() denote matrix trace, H denotes conjugate transpose and
G(z) is the IO dynamics. norm measures the expected root
mean-square value of the output of a system in response to white
noise excitation. Similarly, for noise dynamics,
2 = 1 2 .times. .pi. .times. .intg. .infin. - .infin. .times. Tr (
H .function. ( e j .times. w ) H .times. H .function. ( e j .times.
w ) .times. d .times. .omega. . ##EQU00033##
[0138] Methods and Apparatus, Further Examples. In summary of the
foregoing, FIG. 9 shows aspects of a method 900 for time-efficient
identification of a brain network input-output (IO) dynamics model
for brain stimulation may include, at 910, generating an input
stochastic-switched noise-modulated waveform characterized by at
least one parameter modulated according to a stochastic-switched
noise sequence. The method may further include, at 920, inputting
the input stochastic-switched noise-modulated waveform to a
clinical brain-response system. The method may further include, at
930, recording one or more time-correlated outputs of the clinical
brain-response system responsive to the input stochastic-switched
noise-modulated waveform. The method may further include, at 940,
identifying, as a brain network IO dynamics model for brain
stimulation, a mathematical model that optimally correlates the
input stochastic-switched noise-modulated waveform to the one or
more time-delimited outputs of the clinical brain-response
system.
[0139] In a further aspect, the stochastic-switched noise-modulated
waveform is generalized by a tunable parameter based on a ratio of
a minimum switching time (T.sub.sw) to a time constant (T.sub.c)
for the brain network IO model. For example, the tunable parameter
may be determined by 1-T.sub.sw/T.sub.c in equation 4. Estimating a
value of T.sub.c may be performed, for example, by applying an
on-off input pattern to the clinical brain-response system in open
loop mode and measuring a rise time for achieving a defined initial
portion of a transition period. In an aspect, the defined initial
portion may defined by 1/e (36.79%). The estimating may include
repeating the applying and the measuring, thereby obtaining
multiple values of the rise time, and averaging the multiple values
to obtain the value of T.sub.c.
[0140] In a further aspect, the mathematical model is a linear
state-space model. Identifying the model may include, for example,
constructing a linear state-space model and determining parameters
that most closely fit the observed responses of the clinical system
over a specified range of inputs. In an aspect, the clinical
brain-response system may be an electronic system that simulates
brain response with electrodes, for example, a Neuro Omega.TM.
system as described in connection with FIG. 3A, or a Monte Carlo
simulator. In an alternative aspect, the clinical brain-response
system may include an electrode implanted in a living brain with
the outputs measured by an electrode array, sometimes called a true
brain system.
[0141] In a further aspect, the stochastic-switched noise-modulated
waveform is a binary waveform. The at least one parameter may be
selected from frequency, amplitude, or pulse width. Examples are
provided in connection with FIGS. 2A-D, and elsewhere herein. As
used herein, "stochastic switched" means that the value of each
waveform parameter (e.g., frequency and amplitude) is switched
between two or more values at randomly determined intervals. If the
switching is between two predetermined values only, the waveform is
referred to herein as "binary."
[0142] FIG. 10 is a conceptual block diagram illustrating
components of an apparatus or system 1000 for time-efficient
identification of a brain network input-output (IO) dynamics model
for brain stimulation, as described herein. The apparatus or system
1000 may include additional or more detailed components for
performing functions or process operations as described herein. For
example, the processor 1010 and memory 1016 may contain an
instantiation of a process for identifying an I/O model as
described herein above. As depicted, the apparatus or system 1000
may include functional blocks that can represent functions
implemented by a processor, software, or combination thereof (e.g.,
firmware).
[0143] As illustrated in FIG. 10, the apparatus or system 1000 may
comprise an electrical component 1002 for generating an input
stochastic-switched noise-modulated waveform characterized by at
least one parameter modulated according to a stochastic-switched
noise sequence. The component 1002 may be, or may include, a means
for said generating. Said means may include the processor 1010
coupled to the memory 1016, to an electrical output 1013 connected
to an input of a clinical brain system and to an input 1012
connected to an output of the brain system, the processor executing
an algorithm based on program instructions stored in the memory.
Such algorithm may include a sequence of more detailed operations,
for example, receiving DBS parameters for an input signal (e.g.,
parameters for BN or GBN input) in a digital data format and
generating an analog electrical signal for an electrode array using
an D/A signal generator so as to comply with the DBS
parameters.
[0144] The apparatus 1000 may further include an electrical
component 1003 for inputting the input stochastic-switched
noise-modulated waveform to the clinical brain-response system. The
component 1003 may be, or may include, a means for said inputting.
Said means may include the processor 1010 coupled to the memory
1016, the processor executing an algorithm based on program
instructions stored in the memory. Such algorithm may include a
sequence of more detailed operations, for example, receiving a
sequence of digital data representing binary numbers, converting
the sequence to a waveform using a digital-to-analog converter, and
transmitting the waveform from the output 1013.
[0145] The apparatus 1000 may further include an electrical
component 1004 for recording one or more time-correlated outputs of
the clinical brain-response system responsive to the input
stochastic-switched noise-modulated waveform. The component 1004
may be, or may include, a means for said recording. Said means may
include the processor 1010 coupled to the memory 1016 and to the
input 1012, the processor executing an algorithm based on program
instructions stored in the memory. Such algorithm may include a
sequence of more detailed operations, for example, receiving one or
more analog signals, converting the signals to digital number
sequences using an analog-to-digital converter, and storing the
sequences in a memory correlated by a time index to each other and
to the output sequence.
[0146] The apparatus 1000 may further include an electrical
component 1005 for identifying, as a brain network IO dynamics
model for brain stimulation, a mathematical model that optimally
correlates the input stochastic-switched noise-modulated waveform
to the one or more time-delimited outputs of the clinical
brain-response system. The component 1005 may be, or may include, a
means for said identifying. Said means may include the processor
1010 coupled to the memory 1016, the processor executing an
algorithm based on program instructions stored in the memory. Such
algorithm may include a sequence of more detailed computational
operations, for example, as described in connection with FIGS. 1A-B
and 3D for system identification.
[0147] The apparatus 1000 may optionally include a processor module
1010 having at least one processor. The processor 1010 may be in
operative communication with the modules 1002-1005 via a bus 1010
or similar communication coupling. In the alternative, one or more
of the modules may be instantiated as functional modules in a
memory of the processor. The processor 1010 may affect initiation
and scheduling of the processes or functions performed by
electrical components 1002-1005.
[0148] In related aspects, the apparatus 1000 may include a network
interface module (not shown) operable for communicating with system
components over a computer network. A network interface module may
be, or may include, for example, an Ethernet port or serial port
(e.g., a Universal Serial Bus (USB) port). In further related
aspects, the apparatus 1000 may optionally include a module for
storing information, such as, for example, a memory device 1016.
The computer readable medium or the memory module 1016 may be
operatively coupled to the other components of the apparatus 1000
via the bus 1010 or the like. The memory module 1016 may be adapted
to store computer readable instructions and data for affecting the
processes and behavior of the modules 1002-1005, and subcomponents
thereof, or the processor 1010, or the method 900 and one or more
of the additional operations disclosed herein. The memory module
1016 may retain instructions for executing functions associated
with the modules 1002-1005. While shown as being external to the
memory 1016, it is to be understood that the modules 1002-1005 can
exist within the memory 1016 or an on-chip memory of the processor
1010.
[0149] The apparatus 1000 may include a transceiver (not shown)
configured as a wireless transmitter/receiver, or a wired
transmitter/receiver, for transmitting and receiving a
communication signal to/from another system component such as, for
example, a server or client device. In alternative embodiments, the
processor 1010 may include networked microprocessors from devices
operating over a computer network. In addition, the apparatus 1000
may include any suitable display output and user interface device
for interaction with a user.
[0150] Referring to FIG. 11, a method 1100 for causing a desired
brain response to an input electrical signal may include, at 1110,
providing an input electrical signal to a true brain system via at
least one input electrode. The method 1100 may further include, at
1120, detecting a response of the true brain system to the input
electrical signal via at least one output electrode. The method may
further include, at 1130, comparing the response to a desired brain
response. The method 1100 may further include, at 1140, modulating
the input electrical signal, according to a closed-loop control
algorithm based on a brain network IO dynamics model for brain
stimulation developed by correlating a stochastic-switched
noise-modulated waveform input to the true brain system output,
until the response matches the desired brain response.
[0151] In an aspect of the method 1100, the brain network IO
dynamics model optimally correlates the input stochastic-switched
noise-modulated waveform to the true brain system output, as
described elsewhere herein. In one use case of the method 1100, the
desired brain response mitigates a neuropathic condition, for
example, Parkinson's disease, or depression. In another use case,
the desired brain response heightens a brain performance parameter,
for example, wakefulness, focus, or memory.
[0152] FIG. 12 is a conceptual block diagram illustrating
components of an apparatus or system 1200 for causing a desired
brain response to an input electrical signal. The apparatus or
system 1200 may include additional or more detailed components for
performing functions or process operations as described herein. As
depicted, the apparatus or system 1200 may include functional
blocks that can represent functions implemented by a processor,
software, or combination thereof (e.g., firmware).
[0153] As illustrated in FIG. 12, the apparatus or system 1200 may
comprise an electrical component 1202 for providing an input
electrical signal to a true brain system via at least one input
electrode. The component 1202 may be, or may include, a means for
said providing. Said means may include the processor 1210 coupled
to the memory 1216, and to an output of a sensor array (not shown),
the processor executing an algorithm based on program instructions
stored in the memory. Such algorithm may include a sequence of more
detailed operations, for example, generating an initial waveform as
a digital sequence, converting the waveform to an analog signal,
correlating the output to a time index, and transmitting the analog
signal from a output electrode 1213.
[0154] The apparatus 1200 may further include an electrical
component 1203 for detecting a response of the true brain system to
the input electrical signal via at least one output electrode. The
component 1203 may be, or may include, a means for said detecting.
Said means may include the processor 1210 coupled to the memory
1216, the processor executing an algorithm based on program
instructions stored in the memory. Such algorithm may include a
sequence of more detailed operations, for example, receiving an
analog signal at an input 1212, correlating the input signal to a
time index, and converting the signal to a digital sequence.
[0155] The apparatus 1200 may further include an electrical
component 1204 for comparing the response to a desired brain
response. The component 1204 may be, or may include, a means for
said comparing. Said means may include the processor 1210 coupled
to the memory 1216 and to a sensor (not shown), the processor
executing an algorithm based on program instructions stored in the
memory. Such algorithm may include a sequence of more detailed
operations, for example, retrieving the time-indexed output and
input sequences, and computing a difference over each correlated
time interval.
[0156] The apparatus 1200 may further include an electrical
component 1205 for modulating the input electrical signal,
according to a closed-loop control algorithm based on a brain
network IO dynamics model for brain stimulation developed by
correlating a stochastic-switched noise-modulated waveform input to
the true brain system output, until the response matches the
desired brain response. The component 1205 may be, or may include,
a means for said modulating. Said means may include the processor
1210 coupled to the memory 1216, the processor executing an
algorithm based on program instructions stored in the memory. Such
algorithm may include a sequence of more detailed operations, for
example, applying a closed-loop control scheme based on an
identified brain I/O system model as described more fully elsewhere
herein.
[0157] The apparatus 1200 may optionally include a processor module
1210 having at least one processor. The processor 1210 may be in
operative communication with the modules 1202-1205 via a bus 1212
or similar communication coupling. In the alternative, one or more
of the modules may be instantiated as functional modules in a
memory of the processor. The processor 1210 may affect initiation
and scheduling of the processes or functions performed by
electrical components 1202-1205.
[0158] In related aspects, the apparatus 1200 may include a network
interface module (not shown) operable for communicating with system
components over a computer network. A network interface module may
be, or may include, for example, an Ethernet port or serial port
(e.g., a Universal Serial Bus (USB) port). In further related
aspects, the apparatus 1200 may optionally include a module for
storing information, such as, for example, a memory device 1216.
The computer readable medium or the memory module 1216 may be
operatively coupled to the other components of the apparatus 1200
via the bus 1212 or the like. The memory module 1216 may be adapted
to store computer readable instructions and data for affecting the
processes and behavior of the modules 1202-1205, and subcomponents
thereof, or the processor 1210, or the method 1100 and one or more
of the additional operations disclosed herein. The memory module
1216 may retain instructions for executing functions associated
with the modules 1202-1205. While shown as being external to the
memory 1216, it is to be understood that the modules 1202-1205 can
exist within the memory 1216 or an on-chip memory of the processor
1210.
[0159] The apparatus 1200 may include a transceiver (not shown)
configured as a wireless transmitter/receiver, or a wired
transmitter/receiver, for transmitting and receiving a
communication signal to/from another system component such as, for
example, a server or client device. In alternative embodiments, the
processor 1210 may include networked microprocessors from devices
operating over a computer network. In addition, the apparatus 1200
may include any suitable display output and user interface device
for interaction with a user.
[0160] Those of skill would further appreciate that the various
illustrative logical blocks, modules, circuits, and algorithm steps
described in connection with the aspects disclosed herein may be
implemented as electronic hardware, computer software, or
combinations of both. To clearly illustrate this interchangeability
of hardware and software, various illustrative components, blocks,
modules, circuits, and steps have been described above generally in
terms of their functionality. Whether such functionality is
implemented as hardware or software depends upon the particular
application and design constraints imposed on the overall system.
Skilled artisans may implement the described functionality in
varying ways for each particular application, but such
implementation decisions should not be interpreted as causing a
departure from the scope of the present disclosure.
[0161] As used in this application, the terms "component",
"module", "system", and the like are intended to refer to a
computer-related entity, either hardware, a combination of hardware
and software, software, or software in execution. For example, a
component or a module may be, but are not limited to being, a
process running on a processor, a processor, an object, an
executable, a thread of execution, a program, and/or a computer. By
way of illustration, both an application running on a server and
the server can be a component or a module. One or more components
or modules may reside within a process and/or thread of execution
and a component or module may be localized on one computer and/or
distributed between two or more computers.
[0162] Various aspects will be presented in terms of systems that
may include a number of components, modules, and the like. It is to
be understood and appreciated that the various systems may include
additional components, modules, etc. and/or may not include all of
the components, modules, etc. discussed in connection with the
figures. A combination of these approaches may also be used. The
various aspects disclosed herein can be performed on electrical
devices including devices that utilize touch screen display
technologies, heads-up user interfaces, wearable interfaces, and/or
mouse-and-keyboard type interfaces. Examples of such devices
include VR output devices (e.g., VR headsets), AR output devices
(e.g., AR headsets), computers (desktop and mobile), smart phones,
personal digital assistants (PDAs), and other electronic devices
both wired and wireless.
[0163] In addition, the various illustrative logical blocks,
modules, and circuits described in connection with the aspects
disclosed herein may be implemented or performed with a general
purpose processor, a digital signal processor (DSP), an application
specific integrated circuit (ASIC), a field programmable gate array
(FPGA) or other programmable logic device, discrete gate or
transistor logic, discrete hardware components, or any combination
thereof designed to perform the functions described herein. A
general purpose processor may be a microprocessor, but in the
alternative, the processor may be any conventional processor,
controller, microcontroller, or state machine. A processor may also
be implemented as a combination of computing devices, e.g., a
combination of a DSP and a microprocessor, a plurality of
microprocessors, one or more microprocessors in conjunction with a
DSP core, or any other such configuration.
[0164] Operational aspects disclosed herein may be embodied
directly in hardware, in a software module executed by a processor,
or in a combination of the two. A software module may reside in RAM
memory, flash memory, ROM memory, EPROM memory, EEPROM memory,
registers, hard disk, a removable disk, a CD-ROM, digital versatile
disk (DVD), Blu-ray.TM., or any other form of storage medium known
in the art. An exemplary storage medium is coupled to the processor
such the processor can read information from, and write information
to, the storage medium. In the alternative, the storage medium may
be integral to the processor. The processor and the storage medium
may reside in an ASIC. The ASIC may reside in a client device or
server. In the alternative, the processor and the storage medium
may reside as discrete components in a client device or server.
[0165] Furthermore, the one or more versions may be implemented as
a method, apparatus, or article of manufacture using known
programming and/or engineering techniques to produce software,
firmware, hardware, or any combination thereof to control a
computer to implement the disclosed aspects. Non-transitory
computer readable media can include but are not limited to magnetic
storage devices (e.g., hard disk, floppy disk, magnetic strips, or
other format), optical disks (e.g., compact disk (CD), DVD,
Blu-ray.TM. or other format), smart cards, and flash memory devices
(e.g., card, stick, or other format). Of course, those skilled in
the art will recognize many modifications may be made to this
configuration without departing from the scope of the disclosed
aspects.
[0166] In view of the exemplary systems described supra, methods
that may be implemented in accordance with the disclosed subject
matter have been described with reference to several flow diagrams.
While for purposes of simplicity of explanation, the methods are
shown and described as a series of blocks, it is to be understood
and appreciated that the claimed subject matter is not limited by
the order of the blocks, as some blocks may occur in different
orders and/or concurrently with other blocks from what is depicted
and described herein. Moreover, not all illustrated blocks may be
required to implement the methods described herein. Additionally,
it should be further appreciated that computer-executable
instructions for the methods disclosed herein are capable of being
stored on an article of manufacture to facilitate transporting and
transferring such instructions to computers, which by executing the
instructions are caused to perform the methods.
[0167] Conclusions. A novel computational framework is used to
model and identify input-output dynamics of brain networks in
response to electrical stimulation, with the goal of designing
closed-loop controlled DBS systems. A new clinical simulation
testbed using the Neuro Omega.TM. microelectrode recording and
stimulation system to validate the framework is also developed. The
computational framework consists of a novel input DBS/stimulation
waveform-a pulse train with binary noise (BN)-modulated parameters
to accurately identify the input-output dynamics with a linear
state-space model. This input is extended to generalized binary
noise (GBN)-modulated pulse trains to enable time-efficient
identification; this is done by further estimating the
time-constant of the underlying brain network using on-off
stimulation. Performance of the framework was evaluated within the
clinical simulation testbed by performing two sets of experiments
while incorporating noise, unknown disturbances and stimulation
artifacts. First, open-loop identification experiments using the
(G)BN-modulated DBS/stimulation waveform demonstrated that the new
input DBS/stimulation waveform was critical in accurate model
identification. Second, closed-loop control experiments using the
(G)BN-identified models demonstrated that the BN-identification
framework was essential for accurate closed-loop control; in this
case the treatment of depression serves as a non-limiting example.
Results and methods have significant implications for the
development of closed-loop DBS/stimulation systems or BMIs for
future stimulation-based treatments of a wide range of neurological
disorders including various neuropsychiatric disorders.
[0168] Having thus described embodiments of methods and apparatus
for making use of a stochastic-switched waveform for identification
of I/O brain network dynamics, it should also be appreciated that
various modifications, adaptations, and alternative embodiments
thereof may be made within the scope and spirit of the present
invention. For example, use of a binary noise signal has been
disclosed, but the inventive concepts described above would be
equally applicable to implementations using other
stochastically-switched waveforms that are not binary.
* * * * *