ECE 539Project ReportFall 2001An ANN Approach to EEG Data ScoringAnand LakshmananIntroduction:The human brain is the most complex information-processing structureknown to science. The brain contains some 100 billion neurons which operate by generating and passing electrical signals. The summation ofall this electrical activity results in signals that can be detected and recorded outside the brain. In analogy to the recording of the activity of the heart in an electrocardiogram (ECG), the recording of the brains activity is called an electroencephalogram (EEG).Electroencephalograms, or EEGs are weak electrical signals obtained from electrodes placed on a person’s head. These brain wave signals represent the state of cell activity in the brain, and their interpretation is a major analytical problem.Over the years, physicians and scientists have correlated certain waveforms with the level of an individual’s consciousness, with brain damage which might be present or with certain kinds of brain ailments.The EEG represents dynamics of electrical brain activity on a time scale of milliseconds. Neural systems are capable of generating complex EEG signals with highly nonlinear dynamicsNeural Networks have evolved from the way neurons in the human brain function.Naturally, it makes sense to apply neural networks to aid the analysis of EEGdata that is collected from the human brain!!In psychology related experiments, it is a common practice to compute the total power spectral density(psd) of eeg signals in order to make some important decisions, for example , to classify a human subject as a mentally depressed person or not. In doing so, eye artifacts cause large scale localized errors in the eeg output. Types of eye movements (artifacts)1. horizontal eyeball movements2. vertical eyeball movements3. blinksSo it is a common practice to “score-out” the eye artifact time epochs from eeg data and compute the psd from the rest of the clean eeg data.Thousands of human hours are spent in the classification between normal eeg and an eye artifact data. If ANN approach is applied in this classification problem, it would eliminatethe need for manual scoring and thus would save time and effort in a massive scale.The following paper discusses a pattern classification approach to differentiate between an eye-artifact and a normal eeg signal. A multilayer perceptron network was used to classify the data based on its features. The simplified model used here is startle under noise. Startle refers to an eye blink. Good classification rates are achieved by the networkhowever it is a long way to go before we can surpass other non-linear effects that come into picture which is discussed in the limitations section.Theory:EEG is normally used to record the brain wave in medical treatment.The recording isusually taken by electrodes (small metallic discs) pasted by anelectricity conducting gel to the surface of the scalp. In EEG recording, a powerful electronic amplifier increases severalhundreds or thousands of times the amplitude of the weak signal (lessthan a few micro volts) which is generated in this place. In the past, adevice called galvanometer, which has a pen attached to its pointer,writes on the paper strip, which moves continuously at a fixed speedpast it. In the present time, with the advent of powerful electroniccomputer and very high storage, we can use A/D device to transformsignal between electrode and computer. A lot of data can be recordedand easily analyzed and printed. One pair of electrodes usually makesup a channel. Since earlier times, it is known that the characteristics of EEG activity change in manydifferent situations, particularly with the level of vigilance: alertness, rest, sleep anddreaming. The frequency of wave change can be labeled with names such as alpha, beta, theta and delta. Particular mental tasks also alter the pattern of the waves in different parts of the brain.A small pic to show how eye blink contaminates EEG signalData Collection:Much of the time spent was on data collection.EEG and Startle data are collected on a routine basis at the Psychology Department, UW Madison for various research studies. I was involved in setting up of an experiment where data is collected from human subjects. I sat through some data collection sessions. In addition, a large volume of data was collected and manually scored.Just to give an insight,here is how a typical cap electrode looks likeThis one has only a few electrodes while modern EGI ( Electro Geodesics Inc) have come out with 128 channel net that can directly sit over a subjects head.Startle data was collected using Snapshot Storage software which runs on DOS.The software was obtained from HEM Data Corporation. I wrote a C++ code which was modified from the code used for other studies.This Snap Stream program collects data from channels as per the specified sampling rate and gain settings of an attached Bio-Electric Amplifier.A small dos program controls the gain and filter settings for the different channels.A hardware contour following integrator S7601 from Coulbourn Instruments was used to convert the raw startle to integrated startle data.- This unit has an active amplifier, inverter amplifier and signal mixer in the input to full wave rectify without diode offset error. - The integrator section is a balanced bleed-fill network to maintain equal charge and discharge time constants.- The o/p is the true average of the input signal.- The time constant is adjustable from 50ms to 2 secs making the unit suitable for integration of biopotentials upto the lowest band of EEG signals.The raw and integrated signals look like these:Data is streamed into stimulus files.I used a program startle.m written by Adrian Pederson to read in the data from stimulus file and convert into understandable parameters. Using a probe channel and exciting the channel in occasional intervals , we can cause the blink of subjects which are captured in four epochs as shown in an example.1 2 3 4-50 0 20 120 250-2024681012Time (ms)Voltage (ADC)AR050012 - Startle number 1This one for example is collected for subject 50 trial 12 and displays the 1st among 12 blinks.Epoch 1 is the base line periodEpoch 2 is the wait period for the blink as the probe stimulus has happened.Epoch 3 is the startle capture epoch.( The blink is seen) Epoch 4 is post startle epoch.Amount of data into the neural network:• 43
View Full Document
Unlocking...