Physiological data acquisition and data analytics for neurologists Chou - Ching Lin, MD PhD Department of Neurology & Department of Biomedical Engineering, National Cheng Kung University, Tainan, Taiwan
Physiological data acquisition and data analytics for neurologists
Chou-Ching Lin, MD PhD
Department of Neurology & Department of Biomedical Engineering, National Cheng Kung University, Tainan, Taiwan
2
Introduction
2
Aims of this talk
This lecture is important to provide the dual
understanding on
• what neurologists should know about data acquisition and
basic methods/types of data analysis.
• also to describe to engineers what types of data clinicians
would be looking for in the clinical interpretation and how
current technology can assist in this translation.
3
4
Physiological signals
4
Types of physiological signals
• Single shot Example
– 1D: a number NCV
– 2D: a figure MRS
– 3D: a volume of figures PET, SPECT
• Time series
– 1D: a number EEG (today’s focus)
– 2D: a figure EEG mapping
– 3D: a volume of figures functional MRI
5
Physical nature of physiological signals
• Electrophysiology
• Sonography
• CT
• MRI
• Nuclear radiation
• Movement parameters
• …
6
7
Electrophysiological signals
7
Basic terms in electrophysiology
e- e- e- e- e-
Voltage
Potential
Charge
CurrentResistance
Impedance
V = I RVoltage Current Resistance
8
Membrane potential
• Goldman constant-field equation
V = ln ( )RT Pk·[Ko+] + PNa·[Nao
+] + PCl·[Cli-]
F Pk·[Ki+] + PNa·[Nai
+] + PCl·[Clo-]
[Na+] 15 mM
[K+] 150 mM[Cl-] 10 mM
[Na+] 150 mM[K+] 5 mM
[Cl-] 120 mM
e-e-
e-e-
e-
e+
e+e+
e+e+
Separation of ions leads
to potential field
9
How potential field is generated
• Mainly by separation of
ions by ion pump
– Needs consumption of ATP
• Potential x charge equals to
potential energy
– Can be used to produce ATP
10
Ion channels on cell membrane
Potential
change
11
Voltage Gated Ion Channel
12
Scope of electrophysiological signals
There is at least one electrophysiological study for
an anatomical site interested by Neurologists
• Cerebral cortex EEG
• Major sensory tracts VEP, BAEP, SSEP
• Major motor tract MEP
• Peripheral nerve NCV
• Muscle EMG
13
14
Scheme of signal handling
14
Detection
Collection
Digitization Processing
Analysis
Physiological data Conditioning
Signal flow
15
16
Signal detection
• The nature of the signal
• What kind of device to use
• Impedance match
16
Volume conductor
e- e- e-
e-e-
e-V A / rCharge DistancePotential
17
Properties of volume conductor
though
• Signal amplitude proportional
inversely to distance
to conductivity
• Additive from different sources
e-
e-
All signals in the body will be detected
18
Detection of electrical signal
• Amplitude of EEG ~ 10-6 V
• AAA battery: 1.5 VDifferential amplifier
Scalp
Skull
Cerebral cortex
19
Homogenous
populationWhole brain
Action
potential
Regional
circuit
Level to detect: what technique?
Single cell
Field
potential
EEGMembrane
potential
20
Patch clamp
Intracellular
recording
Extracellular
recording
Probe array
Extracellular
recording
Single probe
Extra-cortical
recording
Multiple probes
Natural filtering by tissues
(E. Niedermeyer, 1999)
Action potentials in nerve fiber
Low pass filtering by tissue
High pass filtering by EEG machine
21
Signals recorded by EEG machine
Heart
Muscle
Noise
(Phone,vibration)
EEG Machine
22
23
Signal conditioning
• To facilitate identification
• To facilitate collection
23
Amplification and sensitivity
1 x
1/2 x
1/4 x
50 (mV)
24
• Amplification is determined by the hardware capability
• Sensitivity is adjustable by the software
Example _Sensitivity
25
LPF: 70 Hz
HPF: 0.3 Hz
Sens: 200 uV p-p
LPF: 70 Hz
HPF: 0.3 Hz
Sens: 100 uV p-p
Signal/Noise ratio (S/N ratio)
• Just make the signal larger won’t make it clearer
• What is more important is S/N ratio
26
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2-4
-2
0
2
4
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2-4
-2
0
2
4
X 1
X 2
27
Signal digitization
• Need to determine the sampling rate
• Need to consider the resolution
Sampling rate
• How precise in horizontal axis
• Usually defined as points in a second (Hz)
28
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2-1.5
-1
-0.5
0
0.5
1
1.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2-1.5
-1
-0.5
0
0.5
1
1.5
100 Hz
5 Hz
Shanon’s theory:
Nyquist frequency
Resolution
• How precise in vertical axis
• Usually defined in bit. – Ex. 4 bits means 16 possible values.
29
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2-1.5
-1
-0.5
0
0.5
1
1.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2-1.5
-1
-0.5
0
0.5
1
1.5
< 0.01
0.25
Matching amplification and resolution
30
0
2
4
6
0
14
0
2
4
6
0
7
0
2
4
6
0
3.5
Range of signal amplitude after amplification
Adequate Too smallToo large
Amplification
Range of signal amplitude after amplification
31
Signal processing
• Filtering
• Phase lock averaging
31
Basic concept of a filter
High pass filter
Pass the fast changing component
32
Low pass filter
Pass the slow changing component
Purpose of using filter
1. To retain certain portion of frequencies
2. Also to remove noise
• Frequency– Unit: Hz (How many swings per second)
– Ex. Heart rate 60/min is equal to ? Hz
33
The concept of domain
• Time domain
• Frequency domain
Time (second)
0 1 0 1
Frequency (Hz)
2 3 4 5 6 7 8 9
34
Fourier Transform
• To switch a signal from time domain to frequency
domain
• Inverse Fourier transform does the contrary
• In fact, there are a lot of transform in Engineering
– A hot example: HHT (Hilbert-Huang Transform)
35
The concept of filter
0 1
Frequency (Hz)
2 3 4 5 6 7 8 9
1
0
0 1 2 3 4 5 6 7 8 9
FT
IFT
0 1
Time (second)0 1
36Time (second) Frequency (Hz)
Types of filter
• Low pass
• High pass
• Band pass
• Band stop
37Frequency
Cut-off frequency
• In real world, the filter cannot be perfect. There is
a gradual transition from pass zone to stop zone.
• Cut-off frequency is defined as the frequency that
the amplitude drop 3 dB.
Frequency
1
-3 dB (~ = 0.7)
Cut-off frequency
38
Low pass filter
0 50 (Hz)
0 10 (Hz)
0 2 (Hz)
0 1 (Hz)
39
High pass filter
0 50 (Hz)
0.4 50 (Hz)
0.8 50 (Hz)
10 50 (Hz)
40
Example _HPF
41
LPF: 70 Hz
HPF: 1.0 Hz
Sens: 200 uV p-p
LPF: 70 Hz
HPF: 0.3 Hz
Sens: 200 uV p-p
Example _LPF
42
LPF: 15 Hz
HPF: 0.3 Hz
Sens: 200 uV p-p
LPF: 70 Hz
HPF: 0.3 Hz
Sens: 200 uV p-p
Example _over filter
43
LPF: 70 Hz
HPF: 2 Hz
Sens: 200 uV p-p
LPF: 70 Hz
HPF: 0.3 Hz
Sens: 200 uV p-p
Phase Locked Averaging
• A technique frequently used in EP
• For a measured signal
T = S + N
• Phase locked averaging
mean (T) = mean (S) + mean (N)
but mean (N) tends to 0,
if N is white noise
stimulus
44
Signal / Noise Ratio _Example
0
2
4
0
2
4
0 5 100
2
4
0
2
4
0
2
4
0 5 100
2
4
N = 0.62 N2 = N/5
S = 0.1 S = 0.1
S + N S + N2
45
46
Signal analysis
46
Linear Methods
• Spectral analyses
– Fourier transformation & power spectrum
• Least estimator
• Principal component analysis
• Common spatial processing or filtering
47
Non-linear Methods
• Independent component analysis
• Wavelet decomposition
• Chaotic system
– Dimension estimation
• Hilbert-Huang Transformation (HHT)
• Neural network
48
Neural Networks (NN)
• Shallow NN
– Ensemble learning with stacking
– Multilayer perceptron (MLP)
• Deep NN
– Convolutional NN
– Recurrent NN
– LSTM (Long Short-Term Memory)
– GAN (Generative Adversarial Network)
49
WNN _Example of shallow NN
1st layer
16
neurons
Input
16
neurons
Output
layer
4 wavelet
decomposition
25
6 s
ample
s/ s
ec
Ex
per
imen
t d
ata Identification
Results
2nd layer
16
neurons
50
Training template
EMG
(mV)
EEG
(mV
)
Beeper
(V) Output
target
3
1
-1
10
-10
200
-200
0
2 4 6 8 10
Time(sec)
0
51
Example Results
0~15 sec
15~30 sec
30~45 sec
45~60 sec
60~75 sec
75~90 sec
90~105 sec
105~120 sec
120~135 sec
135~150 sec
EMG
WNN Results
Movement identified
No movement identifiedAccuracy:60%
52
Deep NN _Example
• Augmented reality (AR) through
an intelligent glasses
MS student Ho
Image processing with DNN
• Textboxes
• CRNN
• MobileNetV2
Real world scenario
Training at a PC with GPU cards
Label recognition with DNN
57