Jimmy Abbas

Neuromorphic Design of Smart Prosthetic and

Therapeutic Systems

Complex visual fields

Control of locomotion

Hippocampus - path planning dynamics at a fixed point?

Olfaction

Reward systems

Risk aversion ?

Auditory cues

Attention

Social behavior

Color

Path planning

System state performance

Sensorimotor integration

Opportunities for Neuromorphic Engineers in Medical Rehabilitation

• smart prostheses• artificial limbs• electrical stimulation systems

Smart ProstheticsNovember 9-11, 2006 - Irvine, CA

Goals:

• Enhance the climate for conducting interdisciplinary research and break down related institutional and systemic barriers

• Stimulate new modes of scientific inquiry

• Encourage communication between scientists and public

NAKFI - Themes to date:2008 - Complexity (Nov. 13-15, 2008) 2007 - Aging / Longevity (Nov. 15-18, 2007)

2006 - Smart Prosthetics2005 - Genomics2004 - Designing Nanostructures2003 - Signaling

http://www.keckfutures.org/

Medtronic Corp.

Cochlear Corp.

Alfred Mann Institute,

Cleveland FES Center

Victhom Human Bionics

Motion Control Corp.

Cyberkinetics, Inc.

Prostheses of Today and Tomorrow

Prostheses of the Future

Intelligent

Responsive

Biomimetic

Mechanical Prostheses:Artificial limbsHeart valvesCerebral shunts

Neural prostheses:Hand graspStanding/steppingBladder/bowel controlExerciseMemory/cognitionNeuromotor therapyVisionHearing

Limb lossSpinal cord injuryStrokeBrain injuryParkinson’s DiseaseDeafnessBlindnessMemory impairmentsALSCerebral palsy

Why?

What?

Task Group Topics• Describe a framework for replacing damaged cortical tissue and

fostering circuit integration to restore neurological function.• Build a prosthesis that will grow with a child (such as a heart valve or

cerebral shunt, or a self healing prosthesis).• Develop a smart prosthetic that can learn better and/or faster.• Brain interfacing with materials: recording and stimulation

electrodes.• Refine technologies to create active orthotic devices.• Structural tissue interfaces: enabling and enhancing continual

maintenance and adaptation to mechanical and biologic factors.• Sensory restoration of perception of limb movement and contact.• Design a functional tissue prosthesis.• Create hybrid prostheses that exploit activity-dependent processes.• Can brain control guide or refine limb control?

Task Group 3: Develop a smart prosthesis that can learn better and/or faster.

client prosthesis

facilitator

Task Group 3: Develop a smart prosthesis that can learn better and/or faster.

client prosthesis

facilitatorclient:• representations used in the brain• how to transmit information• mechanisms of plasticity• how to maximally exploit plasticity

Knowledge/technology gaps:

Task Group 3: Develop a smart prosthesis that can learn better and/or faster.

client prosthesis

facilitator

facilitator:• identification of intermediate milestones

(performance and neurophysiological)• individuality of minimal detectable differences• customizing for specific user groups• engage the user

Knowledge/technology gaps:

Task Group 3: Develop a smart prosthetic that can learn better and/or faster.

client prosthesis

facilitator

Knowledge/technology gaps:

prostheses:• biomorphic sensors/actuators for

compatibility w/ neural reps• detecting and communicating user

intent and motor commands• maintain/improve performance

based on dynamic interface• real-time machine learning • redundancy for versatility, efficiency

seamless integration

where are the ‘smarts’?

cooperative learning

co-adaptationmaximizing plasticity

versatility

automated fitting

minimal cognitive demandsease-of-use

neuromorphic systemsbiomimicry

coordination

sensorimotor integration

information transmission/coding

utilizing plasticity autonomous - direct control

self-repairing

high throughput, high fidelity

amazondotcomificationredundancy

Smart ProstheticsNovember 9-11, 2006 - Irvine, CA

Top five activities (from a list of 34) amputees would like to be able to perform with their electric prostheses

1) Type/use a word processor2) Open a door with a knob3) Tie shoelaces4) Use a spoon or fork5) Drink from a glass.

Transradial Users

1) Type/use a word processor2) Cut meat3) Tie shoelaces4) Open a door with a knob5) Use a spoon or fork

Transhumeral Users

Atkins et al., 1996

Surface EMG/ EEG

Implanted EMG

Direct peripheral nerve interface

Electrocorticogram

Recording w/ penetrating spinal/brain electrodes

Non-Invasive

Centrally Invasive

Command Input Source to Prosthesis

Sensory Feedback to User

Visual

Auditory

Cutaneous Stimulation

Peripheral Nerve Stimulation

Recording w/ penetrating spinal/brain electrodes

Non-Invasive

Centrally Invasive

If everyone is thinking alike then someone isn't thinking. - General George S. Patton

Peripheral Nerve

Horch; Utah, ASU

Electrocorticogram

Moran & Leuthardt; Wash. U.Williams; U. Wisconsin

Nicolelis; Duke U.

Donoghue; Brown U.Cyberkinetics, Inc.

Intracortical arrays

Muscle re-innervation

Kuiken; RIC

Where to put the ‘smarts’in smart prostheses?

Endogenoussystem:

Exogenous system:

neural control

neuromorphic control

biological sensors

engineered sensors

muscles/ tendons

biomorphic actuators

Integrating Engineered and Physiological Systems

• biomimicry• physiological engineered• adaptive systems• system integration

Oscar Pistorius

Hugh Herr, MIT

Spinal cord

Control of locomotion

Spinal cord injury

Locomotor retraining

The spinal cord is a conduit for information transfer from the brain.

The spinal cord is a highly sophisticated computational structure that mediates a wide range of physiological functions.

Recruitment modulation

Rate coding

Co-contraction

StiffnessForce-velocity relationship

Reflexes

Contraction dynamicsFatigue

Length-tension

Biarticular muscleSecondary action of a muscle

Length

tension

Key features of neuromotor control that are either mediated by or have implications for spinal cord organization

Pattern Generator (PG)

Spinal Segmental Circuitry (SSC)

SupraspinalCenters

motoneuronoutput

PGmotorcommands

PG reflexmodulation

output

Musculoskeletal System

ascendingoscillatory

drive

descendingPG inputs

descendingmotor

commands

spinalreflexpaths

PGreflexpathways

supraspinalreflex

pathways

Neural organization for locomotor control:

• Anatomical organization

• Reflexes (feedback)

• Pattern generation

• Multi-level structure

most inputs to motoneurons from the brain are indirect (via spinal interneurons)

• spinal cord mediated reflexes

• supraspinal modulation

... ...Musclespindle Muscle

spindleExtensormuscle

Flexormuscle

Descendingpathway

Motor neurons

Ia inhibitoryinterneuron

Ia afferent

Ia afferent

...... ......Musclespindle Muscle

spindleExtensormuscle

Flexormuscle

Descendingpathway

Motor neurons

Ia inhibitoryinterneuron

Ia afferent

Ia afferent

Spinal Central Pattern Generators (CPGs)

• CPG produces basic pattern of locomotion• neuromodulators (5-HT, NE, DA, NMDA)• sensory feedback integrated into CPG operation• CPG activity is affected by training

From Barbeau et al. Brain Res. Rev. 30: 27, 1999

Brain Spinal Cord

Intracellular/extracellular recordings

Brain Spinal Cord

Intracellular/extracellular recordings

The spinal cord is a conduit for information transfer from the brain.

The spinal cord is a highly sophisticated computational structure that mediates a wide range of physiological functions.

Using spinal cord models for locomotor control

Woergoetter

Hybrid Carbon:Silicon system

• neuromorphic aVLSI circuit • real-time closed loop • 1:1 phase coupling

5 sec

2 V VR

C

E

5 sec

2 V VR

C

E

Brain Spinal Cord

Ventral RootRecordings

C

E

L L

E

C

Brain Spinal Cord

Ventral RootRecordings

Brain Spinal Cord

Ventral RootRecordingsVentral RootRecordings

C

E

L L

E

CC

E

L L

E

C

Using spinal cord models for locomotor control

Jung, et al. 2001

250,000 people with SCI in US

‘complete’ vs. ‘incomplete’

cervical - thoracic

Spinal Cord Injury (SCI)

Personal impact:• sensorimotor function• mobility and transfers• cardiovascular function• bladder/bowel• respiration• sexual function• exercise and recreation

Costs of Spinal Cord Injury:• health care costs in $billions/yr• lost productivity• reduced quality of life

Kentucky Agrability

Spinal Cord Injury

Causes of SCI:

(250,000 people with SCI in US)

motor vehicle crashes

falls

acts of violence

sports

other

Spinal Cord Injury: Clinical Outcomes from the Model Systems, Stover, DeLisa, & Whiteneck, Aspen Publishers,1995

Spinal Cord Injury

At the time of injury:• 63% are < 30 years old• 80% are employed or in

school• 90% have not gone beyond

high school• 54% are unmarried

Neurologic level of injury => degree of impairment• thoracic level lesion => paraplegia (46.2%)• cervical level lesion => tetraplegia (52.9%)

Kwon et al.; The Spine Journal 4:451-464, 2004

Primary injury due to mechanical impact

Secondary injury due to inflammation and response cascade

A window of opportunity for neuroprotective interventions

In: Neurobiology of Spinal Cord Injury, Edts.R.J. Kalb and S.M. Strittmatter, Humana Press, Totowa, N.J.

rat thoracic injury

human cervical injury

Intrinsic properties of motoneurons change after SCI

From Bennett et al. J. Neurophysiol. 86: 1955, 2001

Motoneurons atrophy after spinal injury; exercise preserves structure

From Gazula et al. J. Comp. Neurol. 476: 130, 2004

... ...Musclespindle Muscle

spindleExtensormuscle

Flexormuscle

Descendingpathway

Motor neurons

Ia inhibitoryinterneuron

Ia afferent

Ia afferent

...... ......Musclespindle Muscle

spindleExtensormuscle

Flexormuscle

Descendingpathway

Motor neurons

Ia inhibitoryinterneuron

Ia afferent

Ia afferent

SCI affects reflex modulation spasticity

Bladder-sphincter dyssynergybladder hyperreflexia

Yoshimura, Prog. Neurobiol. 1999

www.robomedica.com

Promoting Locomotor Recovery after Incomplete SCI

www.litegait.com

Partial Weight Bearing Therapy (PWBT)

If everyone is thinking alike then someone isn't thinking. - General George S. Patton

Robot-assistedColumbo; (Lokomat)

IntraspinalMicrostimulation

Mushahwar, Horch, Prochazka

Epidural Spinal Cord Stimulation

Herman, HeDimitrievic

Therapist-assistedHarkema

Reflex stimulationField-Fote

Promoting Functional Recovery After SCI

• overhead harness reduces load• treadmill assists with movement• cyclic loading pattern• exploits activity-dependent plasticity

Supplement with Neuromuscular Electrical Stimulationmore complete sensory pattern to spinal cord?

increased (or more functional) plasticity?

Treadmill Training

Neuromuscular Stimulation

Neuromorphic Adaptive Control

surfaceintramuscularintraneuralcuffepimysial

Case Western Reserve Univ.

Neuromuscular Stimulation: electrode types

NeuroTECH

U. Utah

Muscle nonlinearities are exacerbated by recruitment properties and fatigue

• limited input range• steep-slope regions• changes with fatigue

0

10

20

30

40

50

60

70

0 20 40 60 80 100

stimulation level (%)

angl

e (d

eg) pre-trial

after 12 min(300 cycles)

Muscles SkeletonPatternGenerator

adaptation

PatternShaper

feedback

PG/PS Control System

PG/PS control: PG generates cyclic pattern

- structure based on neural modelPS fine tunes the pattern

- single layer adaptive neural networkObjectives:

customize stimulationaccount for nonlinearities, dynamics, etc.adjust for fatigue

d V (t)/dt = (I (V , t) + I (V , t) + I - V (t) / R ) / Ci i i i i i im syn m PM m inj m m m

)(t)-(-i VVm+1

1 = (t)yoimii

2e

i ij ij isyn j syn syn mI (t) = y (t) g (E - V (t))∑

i i iPM NMDA KCaI (t) = I (t)+ I (t)

Membrane dynamics:

Output function (firing rate):

Membrane currents:

PG Neuron Equations

i i i i iNMDA NMDA NMDA i NMDA mI (t) = K g p (t)(E - V (t))

(t)p CVE A - (t))p-(1 CEV A=dt

(t)pdi

))/-((

i

))/-((i imir

iirmi

i

βαβα ee

i i i i iKCa KCa NMDA KCa mI (t) = g [Ca (t)](E - V (t))

](t)Ca[-(t))V-E(K(t)p=dt

](t)Cad[iNMDANMDAmCaNMDANMDANMDAi

iNMDAiiiii

δρ

Burst initiation:

PG Pacemaker Currents

i i iPM NMDA KCaI (t) = I (t)+ I (t)

Burst termination:

Adapted from Brodin, 1991

Es2 Is2 Ia2

I a1 I s1 Es1

Burst initiationMutual inhibitionReciprocal inhibition

Es Exciter of synergistsIs Inhibitor of synergistsIa Inhibitor of antagonists

Inhibitory synapseExcitatory synapse

___

__

PG Circuit

actualdesired

flexorsextensors

time (sec)

angle (deg)

torque(Nm)

PS z(t)

PG y(t)

actualdesired

time (sec)

angle (deg)

torque(Nm)

PS z(t)

PG y(t)

flexorsExtensorsR1 neuron

Reflexes with phase-resetting

Phase-dependent reflexes

actualdesired

flexorsExtensorsR1 neuron

time (sec)

angle (deg)

torque(Nm)

PS z(t)

PG y(t)

Reflexes without phase-resetting

time of disturbance application (% cycle)

Percent reduction

Peak muscle torque

RMS muscle torque

Phase-dependent reflexes

PG model provides:basic movement rhythmphase-resettingphase-dependent reflexescycle period adjustments

Needs:localized learningmore complex movement patternsadditional reflexes

Muscles SkeletonPatternGenerator

adaptation

PatternShaper

feedback

PG/PS Control System

PG/PS control: PG generates cyclic pattern

- structure based on neural modelPS fine tunes the pattern

- single layer adaptive neural networkObjectives:

customize stimulationaccount for nonlinearities, dynamics, etc.adjust for fatigue

PS Neuron Equations

∑=

=m

jtkjytkjwtzk

1)()()(

muscle/skeleton

PG

PS

zy

jx Y

Ydese

Σ

Σ

0 0.2 0.4 0.6 0.8time (sec)

individual PS neuron activity levels, y1j

0

0.5

1

PS Learning

∑ −=Δn

nTtkjyn

tetkjw )(1)()( η

)()()( tytytw prepostη=Δ

)()()( tytytw preaη=Δ

Hebbian learning algorithm:

Heterosynaptic Hebbian learning algorithm:

Time-averaged learning algorithm:

system output (degrees)

0 0.2 0.4 0.6 0.8 1-0.2-0.1

00.10.2weighted PS

neuron activity levels, w1jy1j

0 5 10 15 20-0.2neuron index, j

20 40 60 80 1000

5

10

15RMS error per cycle (degrees)

0 0.2 0.4 0.6 0.8 1-20-10

01020

0 0.2 0.4 0.6 0.8 1-0.4-0.2

00.20.4

PS output, z1(t)

20 40 60 80 1000

5

10

150 0.2 0.4 0.6 0.8 1

-20-10

01020

0 0.2 0.4 0.6 0.8 1-0.4-0.2

00.20.4

0.2

0 0.2 0.4 0.6 0.8 1-0.2-0.1

00.1

0 5 10 15 20-0.2neuron index, j

time (sec) time (sec)

time (sec)

time (sec)

time (sec)

time (sec)

time (sec)

time (sec)

c5%=14 cycles c5%=4 cycles

RMS10=.57° RMS10=.50°

-0.10

0.10.2

PS neuron weights, w1j

0.2

-0.10

0.1

0 0.2 0.4 0.6 0.8 1time (sec)

0 0.2 0.4 0.6 0.8 1time (sec)

individual PS neuron activity levels, y1j

n=10, na=5, η=.002 n=10, na=15, η=.002

0

0.5

1

n(k)

0

0.5

1

Angle (deg)

0 5 10

0

20

40

Time (sec)

90 95 100

0

20

40

Time (sec)

PG/PS Desired

0 10 20 30 40 500

20

40

60

Cycle #

RMS (%)

N=32

PG/PS PD +/-STD

PG/PS Control

Adapting to account for fatigue

0 50 100 150 200 250 3000

10

20

30

40RMS error (%)

OFFOFF OFFON ON ON

Cycle #

Adaptation:

Mean ± 1 STD (n=16)

Thigh Segment

(Hip Angle)(Deg) -15

0

15

-10

10

30Shank

SegmentAngle (Deg)

0102030

Knee JointAngle (Deg)

Fixed Stim01

01

0Stim by PGPS 0

11

47 48 49 50Cycle Number

-15

0

15

-10

10

30

0102030

0 1 2 3Cycle Number

actual

desired

FWR

gluts

first 3 cycles last 3 cycles

quadshams

control of multi-segment movements

-20

-10

0

10

20First 10 cycles

Hip angle (Deg)

-20

-10

0

10

20Last 10 cycles

20 40 60 80 1000

0.2

0.4

0.6

0.8

1

% of cycle

Normalized hams stim

20 40 60 80 1000

0.2

0.4

0.6

0.8

1

% of cycle0 0

20 40 60 80 1000 20 40 60 80 1000

BF

IL

VL

ST

TA

GM

HumanPre-clinical

80

100

120 Hip

60

90

120 Knee

40

85

130 Ankle

40

65

90 Shoulder

80

115

150 Elbow

Fore Limb

1 1.5 2 2.5 3

Foot Fall

time (sec)

H

G

(deg

) F

(deg

)

E

(deg

)

D

(deg

)

C

(deg

)

B

A Hind Limb

HLLFLLHLRFLR

Pre-clinical model development (Jung, et al., ASU)

BFIL

VL

ST

TA

GM

Thota et al. J. Neurotrauma 22(4): 442-465, 2005

Can we control the stepping movement?

Can we promote functional plasticity?

CK200: FES Exercise In Clinic or Home

These projects are being supported by the NIH-NCMRR: 1R43-HD050006-01, 5R44-HD041820-04 to customKYnetics, Inc.

Disclosure: J. Abbas is part-owner and co-founder of customKYnetics, Inc.

Commercial partnership: Commercial partnership: customKYneticscustomKYnetics, Inc. , Inc.

Adaptive Stimulator for Exercise and Rehabilitation

Stimulation-Augmented Exercise and Neuromotor Therapy

Integrating physiological and artificial systems

Endogenoussystem:

Exogenous system:

neural control

neuromorphic control

biological sensors

engineered sensors

muscles/ tendons

biomorphic actuators

• biomimicry

• intact artificial

• system integration

Center for Adaptive Neural Systems

neural prostheses and advanced prosthetic systems

adaptation in neural systems

neuromorphiccontrol systems

biological inspiration

technology to replace lost or missing functionality

technology to promote learning or adaptation

neuro-inspired system design

Jimmy Abbas

Neuromorphic Design of Smart Prosthetic and Therapeutic Systems

FULTONs c h o o l o f e n g i n e e r i n g s c h o o l o f e n g i n e e r i n g

Ira A.

Acknowledgements:NIHNSFFDAUS ArmyASUBarrow Neurological InstituteBanner Good Samaritan Medical Center

JoAnne RiessPankaj KatariaXia ZhangEric HartmanJunli OuJason Gillette, PhDEd Stites, MD, PhDSusan McDowell, MD

Manoshi BhowmikAndrea DowningAlison SitekRanu Jung, PhDNarayanan Krishnamurthi, PhDKen Horch, PhDDevin Jindrich, PhDSharon Crook, PhDBertan Bakkaloglu, PhDBrian Smith, PhDSteve Philips, PhDRichard Herman, MDJohann Samanta, MDPadma Mahant, MDDenise Campagnolo, MDDan Lieberman, MD