ICDR 2006

Implantable Biomimetic Microelectronics as

Neural Prostheses for Lost Cognitive

Function

Theodore W. Berger, Ph.D.

David Packard Professor of Engineering

Professor of Biomedical Engineering and Neuroscience

Director, Center for Neural Engineering

University of Southern California

Classes of Brain Prostheses

Sensory: Artificial systems to transduce physical energy into electrical impulses for the brain, e.g., artificial retina

Motor: Artificial systems to activate or replace paralyzed limbs, e.g., injectable neuro-muscular stimulators

Goal: Develop a biomimetic model of hippocampus to serve as a neural prosthesis for lost cognitive/memory function

Strategy:

1. Biomimetic model/device that mimics signal processing function of hippocampal neurons/circuits

2. Implement model in VLSI for parallelism, rapid computational speed, and miniaturization

3. Multi-site electrode recording/ stimulation arrays to interface biomimetic device with brain

4. Goal: to “by-pass” damaged brain region with biomimetic cognitive function

long-term memory

short-term memory

Clinical Applications for a Hippocampal Cortical Prosthesis

• Brain trauma / head injury (preferential loss of hippocampal hilar neurons)

1.4 million patients: $56B/yr

• Stroke-induced cortical dysfunction (preferential damage to hippocampal CA1)

5.4 million patients: $57B/yr

• Epilepsy (hippocampal CA3 epileptogenic foci)

2.5 million patients: $12B/yr

• Memory disorders associated with dementia and Alzheimer’s disease (preferential cell loss throughout hippocampal formation)

4.5 million patients: $100B/yr

massive loss of hippocampal CA1 pyramidal cells following

an ischemic episode

pyramidal cell layer

Modeling the Transformation of Input Spatio-Temporal Patterns into Output Spatio-Temporal Patterns

r(x, y, t) = G[k(x, y, ), s(x, y, t)]

Stage 1: Replacing a Component of the Hippocampal Neural Circuit with a Biomimetic VLSI Device

- intrinsic circuitry of hippocampus: trisynaptic cascade of dentate-CA3-CA1 subregions

- develop experimentally-based, biomimetic model of the CA3 subregion

- surgically remove CA3 subregion of living hippocampal brain slice

- through neuromorphic, multi-site electrode array, interface VLSI device with brain slice to functionally replace CA3 subregion and replace whole-circuit dynamics

Hippocampal Model of CA3, Implemented in Hardware, Interfaced to a Slice through a Conformal, Multi-Site Planar Electrode

(1) Four-Pulse Input Train to Dentate

DENTATE CA3

CA1

(2) Dentate Output

(3) FPGA Model: CA3

(4) FPGA Simulated CA3 Output

(5) FPGA Input to CA1

(6) CA1 Output

Reconstitution of Hippocampal Trisynaptic Dynamics After Replacement of CA3 with a Biomimetic, Hardware Model

• random impulse train stimulation of dentate

• 1,500 impulses pre / 1,500 impulses post

• range of intervals: 1 msec – 5 sec

• CA1 field EPSP measured as output

• mean NMSE: 17.5%

Pathway to a Hippocampal ProsthesisHippocampal slice Single circuit replacement

Intact hippocampus Multiple circuit replacement

hippocampal slice: single circuit

intact hippocampus: multiple circuits

• develop biomimetic model of damaged hippocampal region

• establish bi-directional communication between biomimetic device and intact hippocampus

• restore whole circuit nonlinear dynamics: appropriate propagation of spatio-temporal patterns of activity through system

Microelectrode Designs (Univ of Kentucky)

10 Current designs

with improved polyimide mask

8 site microelectrodes

50x50 m

1

15x300 m

2

15x300 m

3

10x10 m

4

20x20 m

5

50x50 m

6

50x100 m

7

25x100 m

8

50x150 m

9

25x300 m

R1

50x150 m20x333 m

S1

15x333 m

W4

20x150 m

S2

15x333 m

W1

20x150 m

50x50 m

W3

20x150 m

W2

20x150 m

Original

SR

Hippocampal Ensemble “Memory” Firing Pattern

Hippocampal Spatio-Temporal Coding of Memory in the Behaving Rat

LEVER

LEVER

LEFT

RIGHT

Encoded Sample Lever Position

CA1

CA3

DG

CA3 CA1CA3 CA1

ElectrodeArray

CA1

CA3

DG

CA3 CA1CA3 CA1

HippocampalElectrodeArray

Reward

Nonmatch “Correct” Choice

=

“Delay” 1-30s

Delayed Nonmatch to Sample Task

NM RewardResponse

PresentLever

DNMS Trial

SampleResponse

Delay sec

NP

Modeling the Transformation of Input Spatio-Temporal Patterns into Output Spatio-Temporal Patterns

r(x, y, t) = G[k(x, y, ), s(x, y, t)]

Four patterns

CA3-CA1 Spatio-Temporal Patterns of Hippocampal Population Activity Recorded

During DNMS Learned Behavior

TEMPORAL

CA1

CA3

DG

SEPTALCA3 CA1

MEDIA

L

ELECTRODEARRAY

CA1

CA3

DG

CA3 CA1

19

16 8

GOAL: Predicting

CA1 Spatio-

Temporal Patterns

of Activity Given

CA3 Spatio-

Temporal Patterns

of Activity

Recorded During

Behavior

• Physiologically-plausible model structure

– Post-synaptic potential (U)

– Dendritic integration (K)

– Threshold ()

– Spike-triggered “after potential” (H)

• Stochastic model

– Noise term ()

• Intrinsic neuronal noise

• Unobserved inputs

– K-S validation based on time-rescaling theorem

– Estimation of firing probability (P)

• Maximum likelihood estimation

– Error function: integral of Gaussian function

– Iterative estimation

A Physiologically-Plausible Stochastic Spike Model

...)()()(),,(

)()(),()()()(

1 0 0 0321321

),(3

1 0 02121

),(2

1 0

),(10

1 2 3

1 2

N

n

M M M

nnnni

N

n

M M

nnni

N

n

M

nnii

i

txtxtxk

txtxktxkkty

Volterra Kernel Model

kernel Volterra:

output:

input:

k

y

x

length memory:

inputs ofnumber :

M

N

...)()()(),,(

)()(),()()()(

0 0 03213213

0 021212

01110

1 2 3

1 21

M M M

M MM

txtxtxk

txtxktxkkty

• Single-Input Single-Output Case

• Multiple-Input Multiple-Output Case

Interpretation of First-, Second-, and Third-Order Kernels for Spike-In, Spike-Out Systems

k2cross

Two-Input / Single-Output (including the 2nd order Cross Interactions)

1 1 1

2 2 2

, , ,

, , ,

s s s

s s s

1 1 1

1 2 1 2 3

2 2 2

1 2

M -1

0 1s 1 2s 1 2 1 1 1 2 3s 1 2 3 1 1 1 2 1 3m=0 m m m m m

M -1

1s 2 2s 1 2 2 1 2 2 3s 1 2 3 2 1 2 2 2m=0 m m

u n = k + k m s (n - m)+ k m m s (n - m )s (n - m ) k m m m s (n - m )s (n - m )s (n - m )+

k m s (n - m)+ k m m s (n - m )s (n - m ) k m m m s (n - m )s (n - m )s (

2

,s

1 2 3

1

1 2

3m m m

2s 1 2 1 1 2 2m m

n -m )+

k m m s (n - m )s (n - m )

r(n) = Threshold [u(n)]

k3self

k2self

+

time

Th

resh

old

k0

Output

Model

t1t3 t5

time

t1t3 t5

k1self

ur(n)

Input 1

Input 2

S1(n)

S2(n)

t2 t4

t4t2

Time-Rescaling Theorem and Kolmogorov-Smirnov Test for Model Accuracy

1 2 3 4 5 6

u1 u2 u3 u4 u5 u6

dttPui

i

)(

Time-Rescaling Theorem

If P predicted by the model is correct, spike interval should be transformed into an exponential random variable u with unitary mean.

u can be further transformed into a uniform random variable v on the interval (0, 1).

)(1 iui ev

v can then be tested with Kolmogorov-Smirnov (KS) plot.

Within 95% confidence boundary: Good model.

Out of boundary: Inaccurate model.

First Order Kernel (Linear) Model

31.0

946)log( L

Second Order (Nonlinear) Self-Kernel Model

30.0

874)log( L

Third Order Self-Kernel Model

30.0

867)log( L

Modeling the Contribution of Interneurons

Right brain

Left brain CA3

-1 -0.5 0 0.5 1Time (sec)

0

0.02

0.04

0.06

0.08

0.1

nr_5_2_1

Autocorrelograms, bin = 2 ms

Pro

babi

lity

-2 -1 0 1 2Time (sec)

0

10

20

30

40

nr_5_2_1

Perievent Histograms, reference = A_NONMATCH, bin = 20 ms

Cou

nts/

bin

-2 -1 0 1 2Time (sec)

0

10

20

30

40

50

nr_5_2_1

Perievent Histograms, reference = A_SAMPLES, bin = 20 ms

Cou

nts/

bin

-2 -1 0 1 2Time (sec)

0

20

40

60

nr_5_2_1

Perievent Histograms, reference = B_SAMPLES, bin = 20 ms

Cou

nts/

bin

-2 -1 0 1 2Time (sec)

0

10

20

30

40

50

nr_5_2_1

Perievent Histograms, reference = B_NONMATCH, bin = 20 ms

Cou

nts/

bin

k1 k2 Sample Non-Match

Left

Right

Peri-Event Histograms

Autocorrelogram

interneuron

Multi-Input Multi-Output Stochastic Model

Array of multi-input single-output models

16 CA3 Inputs7 CA1 Outputs

k1 k2

hRecorded CA1 S-TPattern

Predicted CA1 S-TPattern

Output #4

Predicting Hippocampal Spatio-Temporal Activity with a 16-Input, 7-Output Nonlinear Model: Case 1

16 CA3 Inputs7 CA1 Outputs

k1 k2

hRecorded CA1 S-TPattern

Predicted CA1 S-TPattern

Output #4

Predicting Hippocampal Spatio-Temporal Activity with a 16-Input, 7-Output Nonlinear Model: Case 1

WFUHS 16-Channel Stimulator

High-Voltage Boost and Tri-State Circuit

C.

STIM3 Chip Block DiagramA. B.

Triangle Biosystems STIM3 Programmable 16-Channel Stimulator

• 16 channels, programmable

• programmable parameters: delay, frequency, voltage, polarity, sense-line monitoring of actual pulse delivery

• current delivery capacity: 150 A

• aynchronous pulse generation capacity on each channel

Spatio-Temporal Pattern Stimulation of Hippocampus with MI/MO Model Output

Temporal

CA1

8

91

16

CA1

CA3DG

CA3

Med

ial

Late

ral

ArrayElectrode

Ensemble Firing Pattern

Online Stimulation

Online Analysis

Stimulation Pattern

Online Recording

Hampson & Deadwyler 2006, WFUHS

Predicted Firing Pattern