6. Biomedical Signal Processing

Question 1

what are examples of ions that determine the presence of biopotentials

Answer 1

K+, Na+, Ca2+, Cl-

Question 2

what is an eeg

Answer 2

electroencephalogram

recorded from the scalp surface based on neurons in the cerebral cortex

Question 3

what is an emg

Answer 3

electromyogram

spatial-temporal summation of motor unit action potential of all active motor units

placing needle electrodes on body

Question 4

what is an ehg

Answer 4

electrohysterogram

Question 5

what is an erg

Answer 5

electroetinogram

Question 6

what is an egg

Answer 6

electrogastrogram

Question 7

what is a fecg

Answer 7

fetal electrocardiogram

Question 8

why can’t biopotential signals be used directly

Answer 8

raw signals contain noise, which hampers diagnostic info. needs to be processed via signal processing

Question 9

what are some causes of ecg waveform corruption

Answer 9

electrode contact noise
muscle contraction (EMG)
respiration causes baseline drift
instrument noise

Question 10

what are some signal processing methods

Answer 10

least mean square adaptive filtering
wavelet transform

Question 11

what is least mean square adaptive filtering

Answer 11

• assume that main signal noise is corrupted (n2)
• use a reference signal that is correlated with the signal noise (n2)
• minimise [formula here]
• the new filter estimation can now be applied to new data coming in with similar noise
• gradient descent is used to estimate the unknowns

Question 12

what is freq measure in

Answer 12

Hertz (cycles/sec)

Question 13

what is a fourier transform

Answer 13

goal is to tell us how much of each frequency exists in a signal

represents amplitude of a signal over a time domain in 2d space (x= frequency, y = amplitude)

Question 14

why do we need frequency information

Answer 14

• hard to see information e.g. ECG signals across time domain, easier across frequency domain after decomposition
• ECG diagnosis is hard to make in the original time domain signal so it’s transformed to a frequency signal that is now widely-recognised

Question 15

what are the disadvantages of frequency information

Answer 15

they don’t tell us the point in time, only really valid for stationary signals

Question 16

what transform can be used for non-stationary signals

Answer 16

wavelet

Question 17

what is a wavelet transform & what is it used for

Answer 17

• a response to the problem of FT only capturing global frequency information
• decomposes a function into a set of wavelets by scale and location
• used for ECG

Question 18

how is eye movement, blinking, muscle and heart noise contamination of EEGs mitigated

Answer 18

ICA (independent component analysis)

Question 19

what is ica

Answer 19

independent component analysis

separates & removes a wide variety of artefacts from an EEG by linear decomposition

[cocktail party example]

Question 20

what are ica assumptions

Answer 20

• spatially stable mixtures of temporal and cerebral activity (independent)
• summation of potentials arriving from brain scalp and body is linear at electrodes
• propagation delays are negligible (well-synced)

Question 21

discuss the different types of gradient descent

Answer 21

stochastic, batch, mini-batch

batch computes the gradient over the entire dataset

stochastic is over each training sample

mini-batch is smaller collections of training samples (n), and computing the gradient over these

Question 22

discuss choosing the appropriate learning rate

Answer 22

too small - large computation, could take a long time to converge
too big - hinders convergence
medium - can find the global minimum

Question 23

what are some challenges of gradient descent

Answer 23

• hard to chose perfect learning rate
• same learning rate for all parameters
• can miss the optimal solution if there are saddle points

saddle points are when there are multiple dimensions plateauing at the same error e.g. one slopes up the other slopes down. the gradients are all close to 0 here and the GD function can get stuck

learning rate explanation:
1. hyper-parameter denote the user-defined value, like learning rate;

2. learning rate can influence the training process, which will update the parameter (or weights) of network;
3. think you have two feature, height and age. height is 1.6; age is 20. If you update the same step, if the step is 1, then the updated age is 21; updated height is 2.6, which is definitely unreasonable. Similarly, if your data with features of different magnitude, it is better to set a different learning rate for each feature.