Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
D. Markovic / Slide 2
Radio
2
Low-Power Expertise, New Challenges
SDR DFE Rx/TxMIMO BB
Neural-spike DSP
500 μm
10 μm
MTJ
Bottom Electrode
Top Electrode
Technology
Set/Reset Voltage
Device Size
RP and TMR
Die Area
Supply Voltage
65-nm CMOS
0.5V / 1.1V
5.8kΩ / 5%
190 x 60 nm2
23.43 x 13.05 mm2
1.5V
Flexible logic Post-CMOS
Cognitive DSP BB
Energy harvesting Neuroscience and therapy
New dsp
biomed
1.9
44
mm
1.424 mm
AD
C C
h0
(Diff. V
CO
)
AD
C C
h0
Co
un
tGe
n
AD
C C
h1
(Diff. V
CO
)
AD
C C
h1
Co
un
tGe
n
AD
C C
h2
(Diff. V
CO
)
AD
C C
h2
Co
un
tGe
n
AD
C C
h3
(Diff. V
CO
)
AD
C C
h3
Co
un
tGe
n
Te
st A
DC
(Diff. V
CO
)
Te
st A
DC
Co
un
tGe
n
AD
C C
h4
(Diff. V
CO
)
AD
C C
h4
Co
un
tGe
n
AD
C C
h5
(Diff. V
CO
)
AD
C C
h5
Co
un
tGe
n
AD
C C
h6
(Diff. V
CO
)
AD
C C
h6
Co
un
tGe
n
AD
C C
h7
(Diff. V
CO
)
AD
C C
h7
Co
un
tGe
n
Te
st
Str
uctu
res
Te
st
Str
uctu
res
PM
UP
acke
tiza
tio
n
En
gin
e
SP
I
Ma
ste
r
Co
nfig
. R
eg
An
d C
on
tro
ller
Te
st
Cfg
Re
g
IO
PE
1
PE
2
PE
3
PE
4
PE
5
PE
6
PE
7
PE
8
PE
9
PE
10
PE
11
PE
12
PE
13
PE
14
PE
15
PE
16
2.98 mm
2.9
8 m
m
register bank / scheduler
3.16
mm
2.17 mm
Reg. File Bank
128-2048 ptFFT
Hard-outputSphere
Decoder Soft
-ou
tpu
t B
ank
Pre
-pro
c.
Rx DFE0.4 mm2
1.34 mm
1.20
mm
M1
M2
M3
STA+DTA
Power Est.
TestCircuitry
STA+DTA
MW
FFT
Memory Logic
MW+FFT
Level Shifters
18
20
um
1520 um
1.2
6m
m
1.15mm
0.2
mm
Featu
re Ex.
0.6
5m
m
Mem
ory
(PSD
Sensin
g)
P/S + S/P
Mem
ory
(64
-81
92
-pt FFT)
Mixer+C
IC
64
-81
92
-pt
FFT+P
SD Sen
sing
Level Shifter
Ban
d Se
g.
DSP
Logic
DSP
BR
AM
BR
AM
Logic
DSP
Slice L/M
Slice L/M
Slice L/M
Slice L
DSP-48, Slice L, BRAM
Slice L/M
Slice L/M
Slice L/M
64-8kFFT 16-core UDSP
R-C Bank
MPPT
AD
MN
ILR
O
CTSCP
ModeSensing
Native
NM
OS
D. Markovic / Slide 3
Socio-economic impact in US
• Chronic pain (100M people) → $635B / year
• Alzheimer’s (5.3M people) → $220B / year
• Depression (18M)
• Anxiety (3.3M)
• Epilepsy (2.3M)
• Parkinson’s (1M)
3
Brain Disease is a Growing Problem
A subset of patients (1-10%) would qualify for DBS
Effective therapy
>100x treatment gap
Epilepsy
monitoring:
a gateway
to these
indications
Limited therapy
10x treatment gap
>1,000x treatment gap
Extremely limited or no therapy
Can’t treat complex network diseases with old tools
D. Markovic / Slide 4
Low-Precision Tools = Large Treatment Gap
DBS leads today allowcontinent-level access
4
D. Markovic / Slide 5
• Long procedure, manual control
• Chest-based device, very bulky (34cc)
• Low resolution interface technology
5
Existing DBS Device for Parkinson’s Disease
Medtronic Activa SC
500th DBS patient @ UCLA
May 23, 2013
Courtesy: Dr. Nader Pouratian (UCLA)
6.1 x 5.6 x 1cm
34cc
4 ring
contacts
D. Markovic / Slide 6
• 8 channels, record (up to 30 min)
• Rec + stim (not simultaneously)
• Replace battery (2-5 years)
Therapeutic Device for Epilepsy
NeuroPace RNS-300
FDA
2014
6
D. Markovic / Slide 77
Parkinson's
10 M
Epilepsy
70 M
Deafness
360 M
Depression
350 M
Blindness
285 M
Paralysis
17 M
Alzheimer’s
114 M
Temporallobe
Auditorynerve
ACC
Visualcortex
HC
STN
The Burden of Neurological Disease
NA
OCD
138 M
# of patients worldwide
• Pharmacological therapies are limited
• Surgical therapies are even more limited
• Existing surgical therapies <10% market penetration
D. Markovic / Slide 88
Problems: Why Did Other Indications Fail?
• Same probe design in nearly all DBS applications, despite morphological and anatomical differences in target nuclei
• Mostly open-loop stimulation (on 1 or 2 contacts)▪ Continuous 130-Hz stimulation (amplitude, pulse changes)
• Limited sensing (no sensing, or blank during stimulation)▪ Limited understanding of stimulation tissue response
• Limited wireless data (kbps) and power▪ Limited understanding of deep-brain activity
• Bulky devices, long surgeries
D. Markovic / Slide 9
UCLA Medicine and Engineering
Medicine
Engineering
9
D. Markovic / Slide 10
• Alzheimer’s disease (age 65+)▪ Cost of care in 2015: $221B [1]
▪ Apple 2015 revenue: $233B
• Younger population (0-14), US▪ Epilepsy: 5 children per day [2]
▪ TBI: 1 ER visit per minute [3]
A Challenge: Understand Memory
[1] www.alz.org [2] www.epilepsyfoundation.org [3] www.cdc.gov
Negative effects on declarative memory
10
D. Markovic / Slide 11
• Conscious recollection: uniquely human
▪ Episodic (events), semantic (facts)
Perception and Memory Model
SensoryRegister
Encoding
Perception
Short-TermMemory
WorkingMemory
5-9 items
Long-TermMemory
ActiveMemories
Large storage
Learn (save)
Retrieve
Temporary
Storage
Permanent
Storage
0.5s visual
2-4s auditory~30sec
11
D. Markovic / Slide 12
Neuro-
scienceMedicine
Engineering
12
D. Markovic / Slide 13
Studies of Human Memory Today
Epilepsy monitoring technology
• Wall-plugged instruments
• Deep intracranial probes
• Hospital environment
• Limited closed-loop Courtesy: R. Staba (UCLA)
Memory areas
13
D. Markovic / Slide 14
Our focus: hippocampus and entorhinal cortex
Human Memory Circuits
14
D. Markovic / Slide 15
Local field potential
• Local field potential (LFP)▪ Population activity
▪ Analyze frequency bands
15
What to Measure?
Actionpotential
Frequency (Hz)
Am
plit
ud
e (
V)
10k1k1001010.11µ
10µ
100µ
1m
10m
𝜃(3-8Hz)
High-𝛾(80-130Hz)
Also analyze various temporal correlations
• Action potential (AP)▪ Individual neurons
▪ Firing (time, rate)
▪ ~50x higher fsample
D. Markovic / Slide 16
Non-Topographic Organization of Memory Cells
• Adjacent cells encode different concepts
• Response to multiple instances of a concept
• Single units encode specific memories
16Courtesy: Dr. Itzhak Fried (UCLA)
D. Markovic / Slide 17
• Access to deep brain areas▪ Rare clinical situations
• Need both population and single-unit activity▪ Very power hungry
• Need more channels▪ And higher density…
▪ Study specific memories
Why is Engineering Human Memory Hard?
BF probe
(UCLA)
17
D. Markovic / Slide 18
1. Sensing
2. Stimulation
3. Power & Config
4. Data
6. Update
5. Status and Display
Device Channels Spacing AP + LFP Battery Volume
Activa PC+S 2 x 4 1.5 mm N Y (6.3 Ah) 39 cm3
RNS-300 2 x 4 3.5 mm N Y (0.7 Ah) 13 cm3
Proposed 2 x 32 0.3 mm Y N (RF) <1 cm3 18
D. Markovic / Slide 1919
• Cochlear: stable power only at a narrow distance (→ alignment magnet)
• Proposed: stable power delivery up to 3cm (→ no implant magnet)
Distance
sensitive
Distance
tolerant
> 30mW
Proposed
Cochlear
Benefit: no magnet
Cochlear implant
Revolutionizing Wireless Power
D. Markovic / Slide 2020
D. Markovic / Slide 21
• 40 contacts (10x more)
• Multi-scale, high-density
• Same diameter (1.27mm)
21
Our High-Precision Implantable Lead
Our 40-contact probe
Large
defocused
stimulation
contacts
Medium-sized focal
stimulation contacts
Small contacts for
cell-level recording
150µm
30µm
1mm
1mm 150µm 30µm
D. Markovic / Slide 22
• Stimulation: 0.2-0.6mA
▪ 2 distal macro channels
▪ 300µs per phase
▪ 50Hz pulses for 1s
• Micro-channel recording (7.3mm away from the stimulation site)
• Recording: 30kHz sampling
Dynamic Range in Closed-Loop Systems?
Patient 448, REC
Bipolar stimulation
Micro-chrecording
BF probe
This setup uses wall-plugged electronicsto achieve sufficient input dynamic range
22
D. Markovic / Slide 23
~5b of Headroom at Microwatt Power?
1s
Extra
DR
Sign
al a
mp
litu
de
(m
V)
Time (s)
cleaned
signal
signal +artifact
14mV
20mV
6mV
Patient 448
0
Headroom: 20mV / 0.5mV = 40 (32dB) → +5bits
5 10 15 20 25 30 35 40 45–25
–20
–15
–10
–5
0
5
10
15
20
25
spikes
0.5
−0.5m
V
23
D. Markovic / Slide 24
High DR
24
Key Requirements for Implantable Amplifiers
Single-unit front-end
• Pseudo-R nonlinearity distorts Vout
• R<5GΩ & large caps for low HPC
• Sensitive to process and temp
1 High Zin Low HP corner Low power2 3 4
[1] R. Harrison, et al., IEEE J. Solid-State Circuits, June 2003, pp 958–965.
[2] F. Zhang, et al., IEEE Trans. BioCAS, Jan. 2012, pp 344–355.
[3] T. Denison, et al., IEEE J. Solid-State Circuits, Dec. 2007, pp 2934–2945.
[4] Q. Fan, et al., IEEE J. Solid-State Circuits, July 2011, pp 1534–1543.
Vref
Vout
Cin
Vin gm
CL
CF
Chopper-based LFP front-end
• Electrode offset = large DC current due to reduced Zin
• SoA designs: Zin < 30MΩ(need >1GΩ to avoid Coff-chip) Vref
Vout
Cin
Vin gm
CL
CF
f0
f0
f0
Zin
Coff-chip ICVin
D. Markovic / Slide 25
Voltage-to-voltage/current gain: output saturates
25
Addressing the Dynamic Range Issue
+
Desired Signal
1mV
Interferes
100mV
Signal Recovery
Not Possible
Blind-Source
Separation
Signal Recovery
Possible
Conventional
Proposed
Signal recovery not
possible
Voltage-to-phase gain: output does not saturate
AFE
(conventional receive chains are inadequate)
D. Markovic / Slide 26
(suitable for kHz-rate biosignals)
26
Oscillator-based Analog Front End
EEG ECG EMG
< 1µW < 1µW ~ 5µW
Vinp Vinn
Vbp Vbn
Digital Logic
Dout
Configurable to multiple biosignals
Reference DR THD Vpk Zin Vn,rms VDD Power
Denison, JSSC’07 74dB −60dB 2.5mV 8MΩ 1µV 1.8V 2µW
Muller, ISSCC’14 52dB −48dB 0.5mV 28MΩ 1.3µV 0.5V 2.3µW
This Work 91dB −87dB 100mV >1GΩ 2µV 1.2V 3µW
Coff-chip
W. Jiang, V. Hokhikyan, H. Chandrakumar, V. Karkare, D. Marković, ISSCC 2016, pp. 484-485.
D. Markovic / Slide 27
Full-Duplex Recording and Stimulation
1 10 100
1
10
100
1k
10k
Po
we
r p
er
chan
ne
l (μ
W)
Commercial wall-powered
Peak Linear Input (mV)
ImplantThis work
1,000x
10x
27
D. Markovic / Slide 28
Embedded Computational Memory Model
Raw
data
Pre-processing
LFP
spike
Raw
features
Dim. red
Machine learning
Stim. control
Model
features
Shape, time,
freq, space
offline
Biophysical & machine-learning
Spike ActivityLFP Spectum
Correct (View) Spectrogram for RMH
Fre
quency (
Hz)
Time (s)
-4 -2 0 2 4
5
10
15
20
25
30
35
-4
-3
-2
-1
0
1
2
3
4
5Correct (View) Spectrogram for RMH
Fre
quency (
Hz)
Time (s)
-4 -2 0 2 4
5
10
15
20
25
30
35
-4
-3
-2
-1
0
1
2
3
4
5
LFP Phase
-200 -150 -100 -50 0 50 100 150 2000
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Phase (degrees)
Norm
aliz
ed C
ount
Left HP -- Theta Phase 5 ms around Picture Presentation
Correct
Missed
Spike-Phase Coherence
1
2
3
30
210
60
240
90
270
120
300
150
330
180 0
Right Hippocampus Spike Field Coherence at +- 4 ms
• Patient specific: offline learning & online adaptation
• Closed-loop: non-linear adaptive stim-artifact removal28
D. Markovic / Slide 2929
Demo of UCLA NMU
D. Markovic / Slide 30
S. Basir-Kazeruni, S. Vlaski, H. Salami, A.H. Sayed, and D. Marković, IEEE EMBS Conf. on Neural Eng., pp. 186-189, May 2017.
Measurements on adjacent channel(s)
30
Waveform Shape Agnostic Adaptive Stimulation Artifact Rejection
D. Markovic / Slide 3131
True Concurrent Stimulation + Sensing
signal (s) + artifact (a)
cleaned signal (s )
without ASAR with ASAR
1cm3mm
System Approach to Integration & Miniaturization
Companion External (Trial) Device
LEGO-style assembly
IRB-approved for human recordings at UCLA
• Closed-loop stimulation
• Up to 256 channels
• Wireless charging
• Wireless data
Modular Implant Device
23 mm
5 mm
Neuromodulator
64-channel
sensing and
stimulation
Neuromodulator
Depth
array
Surface
array
64 contacts
D. Markovic / Slide 3838
D. Markovic / Slide 3939
D. Markovic / Slide 40
780µW/cm2
Demonstration of Autonomous Energy
Regulated Pout = 645µWΔT=3.5K effective
+ –170mV
Measured VTEH
Acknowledgment: Rick Staba and Joyel Almajano (UCLA Neurology)
in-vivo (3.5K)
fully autonomous
Early feasibility study
• Smallest footprint (by >6x)
• Best power density (by >7x)
Next: anatomical integration
R-C Bank
MPPT
AD
MN
ILR
O
CTSCP
ModeSensing
Native
NM
OS
D. Rozgić and D. Marković, VLSI’15, pp. 278-279.
1.4mm x 1.4mm
40
D. Markovic / Slide 41
The Brain Basis for Group Dynamics
Wireless data link
r1
rk
rN
sNsks1
Animal k:
rk – sense
sk – stim
Battery-less wireless motes
Record neuronal signatures of group dynamics + injury
Multi-channel, multi-animal closed-loop control system
Autism, ADD,Social NS
41
C. Giza, D. Hovda (UCLA)
D. Markovic / Slide 42
Brain Stethoscope: EEG in Your Pocket
with Josef Parvizi and Chris Chafe (Stanford)
Linking Neurology and Computer Music
42
D. Markovic / Slide 43
Seizure or No Seizure?
43
D. Markovic / Slide 44
Hear Brains of Other People?
Technology:dense EEG array + vocal synthesis
downloadBrainTunes
44
D. Markovic / Slide 45
• Facilitate new clinical science▪ Translation to therapy
Revolutionizing Brain Therapies
45
• Network neurotechnology
• Electrical engineering:Signal processing, control, circuits, power, data, …
Neuro-science
Medicine
Engineering