You are on page 1of 19

BMEN E4420: Biomedical Signal Processing and Signal Model Introduction Paul Sajda psajda@columbia.


1 Spr13

Course logistics
Instructor: Teaching Assistant: Class meeting time/place: TA Office hrs: Webpage: Prerequisites: Paul Sajda, Yrjo Hame, W 1:10 pm-3:55pm, ET 253 TBD linear algebra, some programing language (MATLAB), complex analysis.

2 Spr13

Course logistics
Reading: all are handouts Selected Chapters from Wim van Drongelen, Signal Processing for Neuroscientists: An Introduction to the Analysis of Physiological Signals, Academic Press, 2007 Eugene N. Bruce, Biomedical Signal Processing and Signal Modeling, John Wiley & Sons, 2000 Additional selected handouts

3 Spr13

Biomedical Signal Processing Content

We will cover basic principles of signal processing. We will emphasize examples and focus on electrical signals generated by biological systems (e.g. biopotentials). Particular emphasis on neurophysiological signals: (e.g. EEG) We will introduce concepts from: filtering theory statistical processes pattern recognition information theory probabilistic modeling neurophysiology MATLAB
4 Spr13

5 Spr13

Biomedical Signal Processing and Signal Modeling

Biomedical Signal Processing -Signal processing and statistical modeling methods useful when analyzing biomedical signals, e.g.

Electro and Magneto Encephalography Electro Myograms and Cardiograms Circadian rhythms in body temperature Spike trains Speech ...

Property of biomedical signals: non-linear, non-stationary, nonGaussian

6 Spr13

Linear, Stationary, Normal - The stuff biology is not.

Linear transformation y = L[x]: Def:

y ( t )= L [ a x1 ( t )+ b x 2 ( t )]= a L [ x 1 ( t )]+ b L [ x2 ( t )]

Physics often calls for linear combination of signals: Mass, force, energy Concentrations in solutions Electrical and magnetic fields Intensity of incoherent electromagnetic radiation (X-ray, visible light, radio-waves) Amplitude of acoustic signal. ....

7 Spr13

Linear, Stationary, Normal - The stuff biology is not.

Example: We record frontal EEG electrode y(t). It will be contaminated with eye movement artifacts. Assume eye movements generate electrical source signal, x1(t), and some other frontal brain activity gives source, x2(t). Physics tells us that electrical potentials add up linearly:

y ( t )= [ a b ]

where [a b] represent the coupling coefficients for eye muscle and frontal activity respectively.

[ ]
x 1 (t ) x2(t )

y ( t ) a x1 ( t )= b x2 ( t )
Linearity is crucial because given an estimate of x1(t) and a, for example from an electro-oculogram (EOG), we can subtract its influence on y(t):

Linear, Stationary, Normal - The stuff biology is not.

Unfortunately signals are often distorted by nonlinearity.

>> plot([-x(1,:); x(2,:); x(8,:)]')

a x1 x1
>> plot(-x(1,:),x(8,:))

Common problem is limited dynamic range.

>> plot(-x(1,:),x(2,:))

Linear, Stationary, Normal - The stuff biology is not.

Note that non-linearity can often be identified even in a 1D signal by its harmonic distortions.

Harmonic distortion >> specgram(sin(x))


>> specgram(atan(sin(x))

Linear, Stationary, Normal - The stuff biology is not.

Harmonic distortion explained ... For example distortion of quadratic nonlinearity leads to frequency doubling: Cubic leads triple frequencies: General non-linearity contains all orders according to a Taylor expansion:

x ( t )=sin ( t )
1 1 y ( t )= x ( t )=sin ( t )= cos ( 2 t ) 2 2 3 1 3 y ( t )=sin ( t )= sin ( t ) sin ( 3 t ) 4 4
2 2

1 f ( x) y = f ( x )= n n ! x n=1

x =0



Linear, Stationary, Normal - The stuff biology is not.

Often 'normal' distributions are assumed, i.e. Samples are Gaussian distributed. Important because of many nice properties of the Gaussian probability density function (pdf):

x p ( x )= exp 2 2 2

( )


Convolution of Gaussian remains Gaussian Product of Gaussian remains Gaussian Parameters are easy to estimate. Leads to least squares optimization criteria Sums of many random variables converges to Gaussian

>> hist(randn(1000,1))

Linear, Stationary, Normal - The stuff biology is not.

Unfortunately many natural signals are NOT Gaussian.
On the left is an example of tongue electro-myogram (EMG):

However, if we normalize by estimate of the local standard deviation:

s = std(x(i-10:i+10)); x(i) = x(i)/s;

we obtain something that is close to Gaussian.

13 Spr13

Linear, Stationary, Normal - The stuff biology is not.

The property of heteroscedasticity is often used in the context of financial time series. e.g. NY stock exchange index. It states that the signal is short-term Gaussian with time varying standard deviation.

14 Spr13

Linear, Stationary, Normal - The stuff biology is not.

There are many non-stationary signals that can be explained, to a f i rst approximation, as heteroscedastic. In multiple dimensions these signals are also known as spherical invariant random processes.
Image features MEG activity Speech amplitudes NY stock variation

15 Spr13

Linear, Stationary, Normal - The stuff biology is not.

Many natural signals are not stationary, and not normal, and many systems are not linear. Analysis and signal processing is OFTEN EASIER if one can assume stationary, normal signals and linear systems. It is important to identify the nature of the signals and possibly to apply preprocessing to make the assumptions simpler. Non-linearity may be identified simply by looking at scatter plots, or harmonic distortions if a strong oscillation is present (often 60Hz). Non-Gaussian properties can be identified by looking at histograms. We will use cumulants to assess 'normality' quantitatively.

All signal analysis starts by LOOKING AT THE DATA!

16 Spr13

Good News - No final exam! - However will be a midterm. Bad News LOTS of Assignments: 1. MATLAB programing turn in by uploading to courseworks 2. Proofs easy, just to exercise the notation 3. Reading understand the subject and cover gaps 4. Final Project more info later in semester.

17 Spr13

Programming assignments:

Submit single matlab file called by your name, first_last.m, all lower case e.g. paul_sajda.m Your program loads all data required. Assume data files are in current directory. Include 'clear all, close all' at the beginning of the program. Do not user upper case letters for matlab commands, e.g. Use 'axis' instead of 'AXIS'. Collaboration is OK. If you do wish to submit the similar work, please name your partner and be prepared to be quizzed after class. "Similar" submission are easy to spot, in particular if there are mistakes! Suspected undisclosed collaborations will be rejected. The criteria for approving should be clear. If not, please ask in class. Do not take chances by assuming that your work is "sort of correct".


Assignment 1: Reproduce the four figures on slides 9, 10, 12, 13 from the raw data. Use the files eeg.mat and tongemg.mat. Do your best to scale the axes correctly, but they do not need to be perfect. For help on MATLAB run >> demo >> help Useful functions >> lookfor >> whos

In particular, if you are new to matlab, please take the time to run the demo programs which are a very good introduction: basic matrix operations, line plotting, matrix manipulations, 2-D plots, matlab language introduction, axis properties, graphs and matrices, and maybe some of the desktop environment demos.

19 Spr13