Neurological Signal Processing

Introduction
Hans Berger- recorded Electro Encephalogram- pattern of minute changes in the electric
potential of the scalp- from a Human
EEG reflects functional states of the brain such as
(i)Different stages of sleep
(ii)Wakefulness of sleep
(iii)Metabolic disturbances
EEG as a non-invasive diagnostic tool to many neurological disorders such as
(i)epilepsy
(ii) Early diagnosis and localization of brain tumors
(iii)Coma assessment in intensive care
(iv)Definition and assessment of sleep stages
Thus , there is a need for developing a systematic approach to the analysis of clinical EEG
Challenge here is quantifying EEG, methods for signal analysis are needed to recognize and
classify EEG patterns of practical clinical significance
Signal Analysis is used in neurophysiology to identify 3 areas
(i)EEG
(ii)Evoked Potentials (EP’s)
(iii)Electromyograph(EMG)
These signals are usually recorded to obtain relevant information about the
state (physiological, psychological or pathological) based on the underlying
signal.
In general, these information are corrupted by noise as neurophysiological
signals are recorded at rate of several kilobytes per second with large
number of channels. Ie large quantity of data with only fraction of it with
relevant information.
Thus, both data reduction and extraction of relevant data is common goal
in any biomedical signal analysis method to improve the reliability and
accuracy of the process in extracting useful information.
The Brain and its Potentials
Nervous system with its sensors ( temperature, tactile, visual,
auditory , olfactory, pain) is interface to the environment providing
information through our senses relevant to our movements, our
breathing, our heart rate , our digestion.
Brain controls body growth, development and reproduction.
Nervous system with the exception of the most peripheral nerve does
not grow , renew or repair itself as other body tissues.
Imagine brain as a computer whose hardware is essentially fixed from
birth but whose software is continuously changing throughout life as
we learn from experiences
Brain is self maintaining and occupies a volume of 1500cm³.
The rate of information transfer is slow being 10-30 bits per second.
Processing is both serial and parallel involving large groups of cells and
is faster in some situations than any computer Ex: Pattern recognition.
 There will be no errors in the calculations except in case of
pathological changes.
Reliability is high because of built in redundancy in the brain.
Caton in 1875 observed brain waves in 2 forms of brain electrical
activity- (i) Evoked Potential (EP) (ii) spontaneous ongoing electrical
activity EEG.
Evoked Potential (EP) is the electrical activity at the brain’s surface
evoked by a sensory stimulus or event related potential(ERP),it is
measured over the sensory region of the brain corresponding to the
sensory modality being stimulated
Ex: Occipital region for a visual stimulus
Brain has electrical and chemical potentials that enable it to process information and control
movement
Brain activity is made possible by the interconnections of neurons, which form neural pathways
and circuits. These circuits are involved in processing information and controlling movement.
The brain's potential can also be affected by mental indifference or eye movements.
(i) Electrical potentials-

a) Resting membrane potential- The brain's neurons have a resting membrane potential of about
-70 millivolts (mV), meaning the inside of the cell is more negative than the outside.
b) Action potentials- These electrochemical signals are created when neurons fire, and they travel
along the axon to connect with other neurons.
c) Readiness potential- A slowly increasing surface-negative potential that precedes voluntary
movements.

(ii) Chemical potentials-

a) Neurotransmitters: These chemical messengers are released at synapses between neurons, and
they can either excite or inhibit other neurons.
Ex: dopamine, GABA etc
Electrophysiological Origin of Brain waves
Brain –connected to the outside world through the sensory
organs which act as transducers converting physical energy such
as auditory pressure waves or light waves into a series of pulses
which are conveyed along the nerve fibers to the brain.
Amplitude of the sensory stimulus is encoded or represented as
a pulse repetition frequency .
Brain not only responds to sensory stimulus but also “thinks”.
Both activities invoke cognitive processing at the neuronal
level- reflected in EEG
Neuron / Nerve Cell- basic building block of nervous system.
Aggregation of these neurons-arranged in many complicated ways to process
information and to take decisions.
Nerve Cell- membrane that separates two media- cytoplasm(within cell)
and extra cellular fluid with different ionic concentration.
Membrane- polarized at a potential difference of 70-100 mV with inside
negative relative to the outside.
Dendrites- Carry information to the neuron.
Axon- carries information away from the neuron.
Dendritic field (small many fibers)- feed graded slow potentials to the cell
body.
Potentials –either polarized or hyper polarized .
Actions- increase or decrease the probability of cell discharging along its
axon to next cell
Axon – inter connecting cable between cells, pulse is regenerated along
axon like repeaters in telephone lines.
Conduction continues along the neuron to synaptic junctions at either
outer cell bodies or dendrites or other neurons.
Synaptic junctions – maybe 1000 per axon conduct in only one direction.
When nerve impulse arrives at one of these junctions it triggers the
release of a chemical transmitter that ferries effect of nerve pulse across
the junctions by chemicals.
Chemicals –produce graded slow potentials called post synaptic
potentials (PSPs)- maybe 30mV in the recipient cell and dendrites.
potentials maybe excitatory(EPSP) or inhibitory(IPSP).
**Synapse is like a D/A converter , Cell is A/D converter ,axons transmits
spikes /digital signals with repeaters along its route.
Synapse- specialized junction between two neurons where the
electrical activity in one neuron influences activity in the second.
Most synapses occur between axon terminals of one neuron and the
cell body/dendrite of second.
Synapses- also occur between dendrite and dendrite, dendrite and cell
body, axon and axon.
Pre-synaptic neurons- Neurons- conducting information towards
synapses .
Post-synaptic neurons- Neurons- conducting information away
synapses .
Neuronal information processing

Unipolar neuron has only one

process extending from its cell body
Bipolar neuron has two processes
(one axon and one dendrite)
Multipolar neuron has multiple
processes, with one axon and many
dendrites essentially, they are
classified based on the number of
extensions coming from the cell
body. It is the most common type in
the human body.
Post synaptic neuron- have thousands of synaptic junctions on the surface of its
dendrites or cell body so that information from thousands of pre synaptic nerve
converges upon it.
Single motor neuron-in spinal cord receives around 15,000 synaptic endings ;
certain neurons in the brain receive more than 100,000.
Each activated synapse produces small electrical signal either excitatory or
inhibitory in the post synaptic cell.
One of the thousands of synapses from many different pre-synaptic cells-
converge upon single post synaptic cell
Level of excitability of this cell at any moment –how close membrane potential
is to the threshold –depends upon the number of synapses active at any one time
and how many of them are excitatory and how many of them are inhibitory.
If the post synaptic neuron reaches threshold it will generate action potential,
transmitted along its axon to the terminal branches –diverge to influence
excitability of many other cells.
Post synaptic neurons –function as neural integrators as their output reflects the
sum of all incoming bits of information arriving in the form of excitatory and
inhibitory synaptic inputs.
EEG Analysis
To analyze an EEG signal , a computing device has to be used –digital computer.
EEG analysis has basic operations –
(i) Acquisition of data including quality control of data to remove /avoid
artifacts (Inputs)
(ii)Central Processing
(iii)Presentation of results, graphs-numerical analysis (Outputs)
There are different methods for each portion of EEG analysis.
Traditional method is Time and Frequency domain analysis.
Time and Frequency domain analysis
Refer Textbook +class work
Parametric Model
Due to limitations of traditional techniques, new approaches called model-based signal
analysis are used.
These algorithms decide which part of the signal is to be interpreted as noise and which
part reflects the state of the system.
The basic method is the parametric description of an EEG signal.
EEG signal is considered as a realization of a stochastic process, responsible for the
generation of the EEG to be stationary and that the signal can be described by means of a
parametric model (Auto Regressive/Auto Recursive).
It is important to know how electro-encephalographer reads and interprets EEG data to
automate the analysis of EEG for diagnosis. For the development of such a system some
kind of conceptual framework or a structural model is required.
Neurophysiological structural models do not exist; sources and generators of the EEG are
still not known, most frequently applied models in neurophysiology are general non-
specific models that are imported from other sources which is "Phenomenological" model
that is based on the principles of pattern recognition.
 General Scheme of Pattern recognition

Patterns Features Classes

Pattern recognition for EEG

Description-any
EEG additional information Impression-Diagnosis
Feature (neurological Classification
Extraction symptoms of history)

 Here the only source of information is the trained electro-encephalographer (physician).

 The evaluation given by the physician consists of two parts: the first one (referred to as
"description") is purely descriptive/narrative whereas the second one (impressions) contains
the diagnostic conclusions arrived at.
How does the electro-encephalographer extract the relevant features from the EEG? When
reading the EEG physician will start flicking through the pages, seeking out stretches of
abnormal activity and then describe the different kinds of activity he has found; only then will
he proceed to the diagnostic evaluation.
Doctor spends as much time with the usually rare stretches of abnormal activity as with the
background activity, which although pre-dominant, carries only little diagnostic information.
The basic element here is a "wave" the piece of record between two adjacent extrema of the
same sign.
When many similar waves follow each other, they are single wave described by a set of
"grapho elements” described just by their wavelength (i.e. frequency) and amplitude,
sometimes "regularity (ie. width of spectral peak) and only rarely by harmonics.
The description in the spectral domain has proved most useful: most EEG spectra exhibit a
distinctive resonance structure, peaks and valleys, which can be clearly distinguished.
It is the peaks (rhythms) to which most attention must be paid.
 The diagnostically relevant portion of the spectral domains carry special names, namely the
delta (0.5-3.5 Hz), theta (3.5-7.5 Hz), alpha (7.5-12.5 Hz) and beta (>12.5 Hz).
EEG analysis has been approached from two different view points:-

(i) spectral analysis corresponding to background activity. Here the EEG is

supposed to be stationary (or very slowly varying) and spectra are
computed from the fixed signal segments of about 30 s duration. This
method yields good results when the background activity is abnormal.
However, the short-time structure of the EEG is lost in this approach
Ex: Determination of coma depth or the study of influence of medication
on the EEG..

(ii) single-wave analysis, corresponding to abnormal event analysis. Here the

elementary pattern is the single wave which is then described by a set of
grapho-elements. Accordingly, a very large number of features will be
generated which is complicated. This method yields good results when applied
to the detection of particular, rare waveforms, mostly transients.
Phenomenological Model
The definition of a pattern as "any EEG record of 20 min duration" would be too complex to be of
any use.
So, we have to split this pattern into "elementary basic patterns", which is a general principle in
pattern recognition.
Ex: a speech signal has to be segmented into elementary patterns "word" or "phoneme", a hand
written text into "words" and these even into "letters".
Thus, we apply phenomenological EEG model, i.e. a model based on a study of the phenomenon as
it occurs
The assumptions for which are given in terms of two propositions.

 Proposition 1-Segment Structure

The EEG consists of quasi-stationary segments on which transients may be superimposed. Here,
stationary means that within a segment the short-time spectral estimate does not change with time.
Transients on the other hand are short-time non-stationarities that consist of a single wave or two.

 Proposition 2- Feature Structure

A segment is specified by its length and power spectral estimate while a transient by a set of grapho
elements and time of occurrence.
Once the model is established, we have to take into consideration what
requirements an automated system has to satisfy in order to be useful in
routine clinical work.
Towards this we restrict ourselves to the signal processing part of the system,
i.e. segmentation and feature extraction.
In order to be as free as possible in the choice of available methods we choose
as our instrument a digital computer with the stipulation that it must be
possible to implement the system on a small laboratory computer, possibly
with some minor hardware extension.
Additionally, it must be highly reliable and capable of fast processing and
avoid intermediate storage of signals.
The desirable facts include feature extraction in real time and the ability to
accept any EEG occurring in routine clinical work as long as the reading
quality is acceptable.
As far as features are concerned, they must be as few in number, open
to classification analysis and capable of high data reduction with as
little loss of information as possible.
Then is segmentation of the data into stationary , quasi-stationary
stretches while isolating the transients, if any, providing a spectral
description to each of the segments and characterizing the transients in
terms of features such as amplitude, slope, duration, etc.
The reason for spectral description arises from the consideration that
an EEG signal is a realization of a stochastic process which can be
described by the lower order moments of its probability density
function (mean and variance in the time domain and power spectrum
in the frequency domain).
Segmentation is achieved by comparing the power spectra of various
stretches of the EEG as a function of time.
One way of doing it would be to take the absolute difference of the
respective spectra of the stretches and integrate. But most EEG spectra
have a distinctive form and structure, that is, they are characterized by
peaks and valleys.
Therefore, relative changes in them should be sought. Although log
spectra would appear to be a reasonable candidate, it is not acceptable
as it requires excessive computation.
A better course of action would be to pre-whiten the signal (i.e.
flatten the spectrum) and then take the absolute difference (of the
respective spectra). This calls for the use of adaptive (signal
dependent) pre-whitening filter that has its origin in Linear Prediction
Theory.