Quiz 3 Cochlear & ML Flashcards
The three Hjorth parameters are
Fast way to compute characteristic of time-varying signal (mean power, root-mean-square frequency, and root-mean-square frequency spread).
Activity(signal power) = A = a0
Mobility(mean freq of power spectrum) = M = sqrt(a2/a0)
Complexity(change in freq) = C = sqrt(a4/a0)
a0 = variance
a2 = variance of first derivative of signal
a4 = variance of second derivative of signal
Popular in EEG analysis bc they are based on variance so its faster to compute
Simple spatial filtering
Si = signal from channel i
To get electrical potential difference btwn two electrodes
Si,j = Si-Sj
Laplacian Filtering
Second spatial filter method extracts local activity at electrode i by subtracting average activity in 4 orthogonal neighboring electrodes (x)
Si = Si - 1/4 sum(Si,i for all x)
Common average referencing
Enhances the local activity at electrode i by subtracting the average over all electrodes
Si = Si - 1/N sum(Si,i=1 to N)
Principal Component Analysis
Discover underlying statistical variability in the date and reduce dimensionality from L to a much small dimension. Achieves this by finding the direction of max variance in L-dimensional data
Av = lambdav
resulting L distinct eigenvectors are orthonormal (principa component vectors
larger variance (more distance btwn points)
Behavior of Ag/AgCl electrode
Oxidation of silver on the electrode surface to siliver ions in solution at the interface
Ag <> Ag+ + e-
Ag+ ions combine with Cl- ions in solution to form AgCl
Ag+ + Cl- <> AgCl
Electromyography
EMG is skeletal muscle signals, loud and obvious signal to read, signal read across muscle
Signals captures nerves within muscles firing action potentials. Easier to capture than EEG
types: needle, surface foam and gel, gold cup
EMG Circuit
2 Electrodes on muscle > amplifer (Rg and reference) > HPF(remove lower freq noise) > LPF (remove higher freq noise) > Displau
EMG signal capturing
Place electrode on single large skeletal muscle, reference away from monitored muscle. Don’t place on tendons bc they dont contain nerves
Passive vs Active Electrode
Passive: wet or dry, silver or gold cup
Active: dry that are amplified before sending to BCI
Invasive BCI
possible infections due to penetration of blood-brain barrier, encapsulation of electrodes in reactive tissue that degrade signal overtime, damage to brain circuits
What is sound?
- electromagnetic radiation acoustic waves
- same direction of vibration as direction of travel
- waves exert pressure on medium
Characteristic Impedance of Sound Medium
Z = P (acoustic pressure) / ~ ( RMS volume velocity)
I = P~ = P^2/Z
How hearing works?
Outer ear (pinna + outer ear canal) gather sound
Pina: impedance transformations, sound amp, and direction finding
Middle ear connects to outer by tympanic membrane (eardrum) through the ear canal. Houses the ear ossicles - couple sound vibrations from eardrum via oval window into the cochlea
Inner ear cochlea is a fluid filled chamber where fluid movement is converted into nerve action potentials by hair cells.
Sense of directionality
Differential arrival time at outer ear with arrive time and loudness at the auditory complex of the brain
Cochlear Implant Components
Auditory nerve fibers are intact in patients, neurons can be made to fire with electrical stimulation of appropriate strength, duration, and orientation.
Microphone : picks up sound
External sound processor and power supply: filter, select, converts
Transmitting Circuitry: encodes signal and power are transmitted transcutaneously using RF
Receiver-stimulator package
Electrode Array
Cochlear Implant Electrodes
Neurons at diff positions along cochlea respond to diff frequencies. Electrodes at tip of array stimulate lower freq and vice versa.
Maximize number of largely non-overlapping population of . Unavoidable dure to electrode sitting in highly conductive perilymph and are relatively far from their target neurons. Don’t want too many electrodes or frequencies will travel to neighboring electrodes
Intensity of Sound
Determined by number of neurons activated and their firing rate. Dependent on the amplitude of the stimulus current. One can recreate normal neural response if enough independent spectral channels could be excited
Safe Current Density
Nondamaging stimulation by using short duration biphasic current pulses delivered by platinum electrodes. Biphasic pulse ensures the electrochemical reactions that take place are reversible and localized to the electrode tissue interface
Cochlear Implant Signal Processing Summary
Electrode array stimulate neuron response along a short section of basilar membrane. Diff electrodes activated by diff parts of sound spectrum. Range of stimulation is much smaller so processors must compress the range of sound
Compressed analog principal
Using automatic gain control before the signal is passed through a bank of bandpass filters. Analog signals applied to electrodes
Continuous interleaved sampling
Improved compressed analog principal
Apply short impulses to electrodes rather than analog signals to band pass filter. Electrodes are excited sequentially with small time intervals between them in a process CIS. Only one channel opened to prevent overlaps. Maximize method through filter spacing, envelope cutoff frequencies, shape of compression function, and stimulation rate
Envelope Detection
1) lPF with cutoff 200-400Hz
2) Hilbert transform: vowel output by a single BPF
Compression Function
Compression of the envelope transforms the large dynamic range of the acoustic signal into the small dynamic range required by electrical stimulation
Stimulation Rate
Rate at which current pulses are delivered to each electrode (800-2,500 pps)
What to include in an abstract?
Goal, contribution to methods, and how to demonstrate results
What is machine learning?
scientific study of algorithms and statistic models that computer systems use to perform a specific task without explicit instructions and relies on pattern and inference
Data, Pattern, Algorithm, Computation
Classification, vs regression vs unsupervised
Classification: discrete output category
Regression: continuous result (know position)
Unsupervised: interpret data
Why ML for BCI?
Need to find and use complex patter in BCI data. Need to design/apply algorithms to conduct computations
ML work flow
Data aq > data preprocessing > feature extraction > train/build model > test/eval > improve
Basic BCI Circuit Layout
Electrode > 1-100k amp > anti-aliasing filter > ADC > BPF > Notch filter > EEG output
Nyquist sampling theorem
continuous <> discrete
sample rate allows discrete sample to capture all the info in a continuous time signal of finite bandwidth
x(t) is spaces 1/(2F) a part
min sampling freq that will not lose signals info is 2x highest freq
Mu rhythms
repeat at a freq of 7.5-12.5 hz
prominent when body is physically at rest
occur over the motor cortex
form of alpha wave (8-13 hz)
Input and outputs of ML model
Input: feature vectors
Output: classification/clustering results
X1, X2, X3»_space; h(hypothesis)»_space; Y1’,Y2’..
where Y1’~= Y1
What is a feature vector?
feature, attribute, input, etc.
Feature space
You never know if you covered the space completely, hence why you need large sample size.
Binary Classification
Classification > Supervised > learn a function that maps an input too output based on example input-output pairs. Find a boundary btwn two classes based on the labeled training data
Find h based on the Y of X
Linear Discriminant Analysis
Has some similarity to PCA but focuses on maximizing separability among categories
Find a hyper plane, different categories are optimally separated. Means as far a possible and variance as small as possible
Hyper plane = h(X) = 0 = W^TX + W0
assign label 1 or -1
W^TX + W0 >0 then label 1
W^TX + W0 <0 then -1
