EEG ANALYSIS: ANN APPROACH

Transcription

1 EEG ANALYSIS: ANN APPROACH

2 CHAPTER 5 EEG ANALYSIS: ANN APPROACH 5.1 INTRODUCTION The analysis of EEG signals using ANN deals with developing a network in order to establish a relation between input and output neurons. The ANN approach is to determine the features conveying the relevant information for classification and to develop algorithm to classify patterns based on these features. Several sets of features of visual recognition based on amplitude, duration measurement are taken and artificial neural networks are trained using different test sets for various patterns. While selecting the features, the visual recognition criteria are implemented and the disease of a patient is reported, matched with results of network. These features are mainly amplitude and duration measurement taken on significant points on the waveform. The problem of extracting features in an EEG signal was tackled mainly with spectral analysis methods. Energy content of EEG channels in conventional frequency bands adopted by doctors, other techniques include time framework periodicity, auto and cross correlations, mimetic analysis, which extracts template segments from baseline data records. 5.2 CLASSIFICATION BASED ON FREQUENCY Electroencephalography (EEG) waveforms are classified according to their frequency, amplitude and shape as well as the location on the scalp at which they are recorded. The most familiar classification uses EEG waveform based on frequency of a signal [36]. 44

3 Information about signal frequency and shape is combined with the age of the patient, state of alertness or sleep and head site to determine significance. Normal EEG waveforms are defined by the following criteria: Frequency (Hz) is the initial characteristic used to define normal or abnormal EEG rhythms. Most waves of 7.5 Hz and higher frequencies are normal findings in the EEG of an awake adult. Waves with a frequency of 7 Hz or less often are classed as abnormal in awake adults, although they normally can be seen in children or in adults who are a sleep. In certain situations, EEG waveforms of an appropriate frequency for age and state of alertness are considered abnormal because they occur at an inappropriate scalp location or demonstrate irregularities in rhythm or amplitude. Some waves are recognized by their shape, head distribution and symmetry. Certain patterns are normal at specific ages or states of alertness and sleep. The morphology of a wave resemble specific shapes, such as vertex (V) waves seen over the vertex of the scalp in stage 2 sleep or triphasic waves that occur in the setting of various encephalopathies. The EEG signals occur in the frequency range of 0.5 to 30 Hz. The signals are classified depending upon four frequency ranges. This type of classification has a great physiological significance. 45

4 The EEG frequency spectrum is given in table 5.1 as under; Table 5.1 Type of EEG band and frequency range Band Type Delta Theta Alpha Beta I Beta II Frequency Range 0.5 Hz to 4.0 Hz 4.0 Hz to 8.0 Hz 8.0 Hz to 13.0 Hz 13.0 Hz to 22.0 Hz 22.0 Hz to 30.0 Hz Delta Waves ( Hz) Delta waves include all the waves below 4hz.A large amplitude irregular waves occur in deep sleep, in infancy and in serious organic brain disease. The delta can occur in the cortex independently of the activities of the lower region of the brain when these waves are present in the EEG of the adult, which is an abnormal finding. These slow waves have a frequency of 3 Hz or less. They normally are found in deep sleep in adults as well as in infants and children s. Delta waves are abnormal in the awake adult. Delta waves have the largest amplitude of all waves. Delta waves can be focal or diffuse Theta Waves ( Hz) The theta waves lie in the frequency range of 4 to 8 Hz. These occur mainly in the parietal and temporal region in the children, but they also occur in the emotional stress particularly in disappointment and frustration. 46

5 Theta and Delta waves are relatively slow activities generally inconspicuous in the tracings of normal alert subjects. It is common for the central region to exhibit considerable theta activity. Rarely, diffuse activity of 7 to 8 Hz appears as a dominant rhythm. Brief mixture of alpha-theta bursts and single theta or exceptionally delta waves may occur independently in left or right temporal region. Their incidence is greater on the left side. The slow waves are masked by superimposed alpha waves. Single theta and delta waves are present in posterior head regions in some adults. Theta waves are normally found in sleep at any age. In awake adults, these waves are abnormal if they occur in excess. Theta and delta waves collectively are known as slow waves Alpha Waves ( Hz) Alpha waves are found in all age groups but most common in adults. They occur rhythmically on both sides of the head but are often slightly higher in amplitude on the no dominant side, especially in right-handed individuals. They tend to be present posterior more than anteriorly and are especially prominent with closed eyes with relaxation. Alpha activity disappears normally with attention, mental arithmetic, stress, and opening eyes. In most instances, it is regarded as a normal waveform. An abnormal exception is alpha coma, most often caused by hypoxic-isocheims encephalopathy of destructive processes in the pones, intracerebral. In alpha coma, alpha waves are distributed 47

6 uniformly both anteriorly and posteriorly in these patients, who are unresponsive to stimuli. The alpha rhythm arises from posterior portion of the brain. Its amplitude is greatest in either the parietal or the occipital region, the anterior extent varying widely from person to person. The frequency of alpha rhythm is less than 8 Hz in children. Recurrent maximum peak amplitude rarely exceeds 100 pv and usually falls within the range 20 to 60 pv. ' Beta Waves ( Hz) Beta waves are observed in all age groups. They tend to be small in amplitude and usually are symmetric and more evident interiorly. Beta waves normally occur in the frequency range of 14 to 30Hz and sometimes during intense mental activity at high frequency of 50Hz. These are most frequently recorded in the frontal and parietal region of the scalp. They are further classified as betal and beta Beta I ( Hz) The Betal waves have the frequency of 13 to 22Hz, which is nearly twice of alpha waves and effective as much as alpha waves. In addition, they disappear during sleep and are replaced by asynchronous waves Beta II ( Hz) or Gamma Waves The Beta2 or Gamma waves have the frequency range of 22 to 30 Hz. These waves appear during intense activity in central nervous system 48

7 or during tension and have frequency range slightly higher than that of the betal waves. Beta waves may be distinguished by location, but to some extent by patterning and frequency. Under normal conditions, the amplitude of beta activity is low, only exceptionally exceeding 20pv. Although the entire cortex has the capacity to produce waves having frequencies above the alpha range 18 to 24 Hz, rhythm of central region is ordinarily the most prominent of the beta patterns at rest. Another type of beta activity is that which suppliants the alpha rhythm when the latter becomes desynchronized, fig 5.1. DELTA Fig 5.1 beta, alpha, theta, and delta waves 53 SPECIAL WAVEFORMS Some of the waveforms present in the original EEG wave being very useful for diagnosis of patients are treated as special waveforms (a) Spikes Spike is rapid, brief and bi-directional deflection of duration less than 80 ms. A single spike may be monophasic, biphasic or triphasic, additional phases placing the discharges in category of polyspikes or multiple spikes. Fig

8 .\a/&rs'&rshr\*aa wiai/^iaaaa Fig.5.2 waveforms of spikes 5.3.1(b) Polyspike Wave Fig.5.3 polyspike waves Polyspikes is a form of spike wave in which each slow wave is accompanied by two or more spikes. The usual pattern is that the spike wave is faster than 3 Hz, usually 3.5 to 4.5 Hz. It is often associated with seizures. It should not be confused with 6Hz spike, otherwise known as phantom spike. 50

9 5.3.2 Lambda Waves These are the waves, which occur in the posterior head region in association with scanning movements of the eyes. They are identified more readily in children. A surface positive deflection of 100 to 150 msec is the constant feature of the pattern, but the subsequent surface negative deflection that extends its duration to 300 to 400 msec. The amplitude generally doesn t exceed 5pv. In a child, amplitude may exceed 100 to 200 (iv. POSTS are triangular waves that occur in the bilateral occipital regions as positive up going waves. They can be multiple and usually are symmetric. POSTS occur in sleeping patients and said to be most evident in stage 2 of sleep, although they are not uncommon in stage 1. POSTS are similar, if not identical to lambda waves both morphologically and in the occipital distribution. VAAAa/\aAAA/ Fig 5.4 lambda waves and posts Lambda and POSTS are similar morphologically, and have a triangular shape. They occur posteriorly and symmetrically. POSTS stand for positive occipital transients of sleep and occur in stage 2 sleep. 51

10 Lambda occurs in the awake patient when the eyes stare at blank surfaces. Both are normal waveforms and can occur in long or short runs. Lambda waves occur in the occipital regions bilaterally as positive up going. They are triangular and generally symmetric. They occur in the awake patient and to be most evident, when the subject starts at a blank uniform surface. Lambda waves occur when reading and occasionally when watching TV. Morphologically, they are similar to POSTS both in form and in occipital distribution 53.3 Mu Rhythm This activity is like the alpha rhythm. Independent foci of this activity arise from front central and temporal regions. The frequency of the activity is slightly slower than the alpha rhythm but at the higher amplitude. Mu activity is a rhythm in which the waves have a shape suggestive of a wicket fence with sharp tips and rounded bases. It may show phase reversal between two channels. The frequency is generally half of the fast activity present. Mu waves are runs of rhythmic activity that have a specific shape. They are rounded in one direction with a sharp side in the other direction. Frequency is one half of the fast (beta) activity. Mu waves disappear with motor acts of the contra lateral hand or arm. Unlike alpha activity, they are not blocked by eye opening. 52

11 /wwwvw\ \AAAAAAAAA/ ^aaaaaaaaa Fig.5.5: Mu-rhythm waves Spike and Slow Wave This is a special waveform occurring at the time of grandmal epilepsy. Spike and wave format is seen at all ages but most often in children. It consists of a spike, which is probably generated in the cortex, and a large amplitude slow wave usually delta waves, thought to originate from thalamic structures, occurring recurrently. (Fig 5.5) They may occur synchronously and symmetrically in the generalized epilepsies or foeally in the partial ones. In the generalized types of spike and wave, petitmal is characterized by 3 Hz spike-wave, while slow spike wave occurs more usually with brain injury. Continuous Spikes are called as poly spikes; particularly occur at the time of petitmal epilepsy. (Fig. 5.3) K-Complex Waves These waves specifically occur at the time of sleep [37]. These waveforms are occurring periodically. The most peculiar fact about the K-complexes is that they should have the duration more than 0.5 sec. K- complexes are represented by sharp positive wave followed by a sharp negative wave and not easily distinguished from background activity. (Fig 6.1) 53

12 K-complex waves are large amplitude delta frequency waves, sometimes with a sharp apex. They can occur throughout the brain and usually are higher in amplitude and more prominent in the bifrontal regions. Usually symmetric, they occur each time die patient is aroused partially from sleep. Semi arousal often follows brief noises; with longer sounds, repeated K complexes can occur. K complexes sometimes are followed by runs of generalized rhythmic theta waves; die whole complex is termed an arousal burst. Sleep spindle is a short sequence of minomorphic waves having a uniform appearance. (Fig 5.6) 53.6 Sleep Spindles Spindles are groups of waves that occur during many sleep stages but especially in stage II. (Fig 5.7) They have frequencies in the upper levels of alpha or lower levels of beta. Lasting for a second or less, they increase in amplitude initially and then decrease slowly, resembles a spindle. 54

13 Fig. 5.7 Slow sleep spindle in EEG 5.4 FILTERS FOR CLASSIFICATION OF EEG SIGNALS The Digital filters for classification of signals depending on frequency of signals is implemented and designed in MATLAB. The filters specifically use the standard mathematical methods for the analysis of EEG. The purpose of filtering is to isolate and make available subpart of the total or raw EEG signal. Filters are used to pass and display only those frequency components that lie between 4-8Hz, which refer to as theta waves and other EEG bands. The process of filtering, however, distorts the signal in ways other than those intended. The signal is delayed in passing through the filter and arrives at later point in time. Effects such as phase shift, bandwidth can change the appearance of the waveforms. 5.5 BACK PROPAGATION FOR CLASSIFICATION OF EEG SIGNALS The backpropagation model of Neural networks which works excellent for detection of the nonlinear waveforms can be used for detection of specific patterns of the physiological signals of EEG waves. The backpropagation network contains the input layer, output layer and 55

14 hidden layer. The input layer neurons are determined by the length of the vector of input signal. A completely connected, feed-forward ANN used in this research consists of a set of neuron-like 'units' that generate outputs by applying a nonlinear function to a weighted sum of all their inputs. Each of these units belongs to a particular layer of the ANN, takes input from every unit in the layer below it or from the input vector, in the case of the bottom layer and deliver its output to every unit in the layer above it or to the output vector, in the case of the top layer. The weights on all these connections between units of neighboring layers are initially random, so the performance of the network begins at a chance level. During the training of the network, input vectors are presented to the bottom layer and data propagates forward to produce an output vector at the top layer. An error measure is generated based on the difference between the desired output vector and the output vector that was actually produced. Back-propagation, a gradient descent method, is used to propagate a correction for this error backward along the connections between units, altering the weights of these connections. Trained networks are then tested on new data that were not part of the training set [38]. 5.6 NEURAL NETWORK MODELS FOR EEG ANALYSIS Fundamental Architecture This section presents the architecture of the network that is used with the backpropagation algorithm - the multilayer feedforward network. The routines in the Neural Network Toolbox can be used to train more networks that are general. An elementary neuron with R inputs is shown below. Each input is weighted with an appropriate w. The sum of the weighted inputs and the 56

15 bias forms the input to the transfer function f Neurons may use any differentiable transfer function/to generate the output. (Fig. 5.8) Input General Neuron r...*\ Where... H * Number of elements in input vector Fig.5.8 Elementary Neuron Linear transfer function purelin is used in backpropagation networks (Fig. 5.9) n a piinrtln(m) Linear Transfer Function Fig 5.9 LTF in backpropagation If the last layer of a multilayer network has sigmoid neurons, then the outputs of the network are limited to a small range. If linear output neurons are used the network outputs can take on any value. In backpropagation, it is important to calculate the derivatives of any transfer functions used. Each of the transfer functions above, tansig, 57

16 logsig, and purelin, has a corresponding derivative function. To get the name of a transfer function s associated derivative function, call the transfer function with the string 'deriv'. tansig ('deriv') ans = dtansig The three transfer functions are used for backpropagation, but other differentiable transfer functions can be created and used with backpropagation if desired Feedforward Network A single-layer network of S logsig neurons having R inputs as shown below in full detail on the left and with a layer diagram on the right (fig 5.10) Input Layer of Neurons Input r Layer of Neurons l J a= f (Wp t b) = f (Wp b) Where... R = numberof elements in input vector s* numberof neurons in layer Fig: 5.10 feedforward network Feedforward networks have one or more hidden layers of sigmoid neurons followed by an output layer of linear neurons. Multiple layers of neurons with nonlinear transfer functions allow the network to learn nonlinear and linear relationships between input and output vectors. The linear output layer lets the network produce values outside the range -1 to 58

17 +1. On the other hand, to constrain the outputs of a network (such as between 0 and 1), then the output layer should use a sigmoid transfer function (such as logsig). For multiple-layer networks, the number of the layers are used to determine the superscript on the weight matrices. The appropriate notation is used in the two-layer tansig/purelin network. input Hidden Layer Output Layer...\ t... n it linwg rtwi»pt tbu * porrlfiflji-iinluj Fig: 5.11 Neural Network concept for analysis of EEG signal Network can be used as a general function approximation. It can approximate any function with a finite number of discontinuities, arbitrarily well, given sufficient neurons in the hidden layer for 256 input neurons and 1 output neuron Creating a Neural Network for Training EEG Data The first step in training a feedforward network is to create the network object. The function newff in MATLAB creates a feedforward network. It requires 256 inputs and returns the network object. The first input is 64 by 256 matrix of minimum and maximum values for each of the R elements of the input vector. The second input is an array 59

18 containing the sizes of each layer. The third input is a cell array containing the names of the transfer functions to be used in each layer. The final input contains the name of the training function to be used. There is one input vector with 256 elements. The values for the first element of the input vector range between -5.0 and 5.0, the values of the second element of the input vector range between 0 and 5. There are 256 neurons in the first layer and one neuron in the output layer. The transfer function in the first layer is tan-sigmoid, and the output layer transfer function is linear, which creates the network object and initializes the weights and biases of the network; therefore, the network is ready for training. There are times when may want to reinitialize the weights, or to perform a custom initialization. The next section explains the details of the initialization process. Before training a feed forward network, the weights and biases must be initialized. The newff command in MATLAB will automatically initialize the weights, but may want to reinitialize them. This can be done with the command init. This function takes a network object as input and returns a network object with all weights and biases initialized. Once the network weights and biases have been initialized, the network is ready for training. The network can be trained for function approximation (nonlinear regression), pattern association, or pattern classification. The training process requires a set of examples of proper network behavior - network inputs p and target outputs t. During training the weights and biases of the network are iteratively adjusted to minimize the network performance function (net.performfcn.) The default performance function for feedforward networks is mean square error and average squared error between the networks. Several different training algorithms for feed forward networks are used the gradient of the 60

19 performance function to determine how to adjust the weights to minimize performance. The gradient is determined using a technique called backpropagation, which involves performing computations backwards through the network. The backpropagation computation is derived using the chain rule of calculus and is described. In the basic backpropagation training algorithm, the weights are moved in the direction of the negative gradient Backpropagation Algorithm There are many variations of the backpropagation algorithm, of backpropagation learning updates the network weights and biases in the direction in which the performance function decreases most rapidly the negative of the gradient. One iteration of this algorithm can be written where a vector of current weights and biases the current gradient. There are different ways in which this gradient descent algorithm can be implemented: incremental mode and batch mode. In the incremental mode, the gradient is computed and the weights are updated after each input is applied to the network. In the batch mode, all of the inputs are applied to the network before the weights are updated. In batch mode the weights and biases of the network are updated only after the entire training set has been applied to the network. The gradients calculated at each training example are added together to determine the change in the weights and biases [39] Faster Training The testing on several high performance algorithms that can converge from ten to hundred times faster than the algorithms discussed previously. These faster algorithms fall into two main categories. The 61

20 first category uses heuristic techniques, which were developed from an analysis of the performance of the standard steepest descent algorithm. One heuristic modification is the momentum technique and second category of fast algorithms uses standard numerical optimization techniques Variable Learning Rate With standard steepest descent, the learning rate is held constant throughout training. The performance of the algorithm is very sensitive to the proper setting of the learning rate. If the learning rate is set too high, the algorithm may oscillate and become unstable. If the learning rate is too small, the algorithm will take too long to converge. It is not practical to determine the optimal setting for the learning rate before training and optimal learning rate changes during the training process, as the algorithm moves across the performance surface. The performance of the steepest descent algorithm can be improved, if we allow the learning rate to change during the training process. An adaptive learning rate will attempt to keep the learning step size as large as possible while keeping learning stable. The learning rate is made responsive to the complexity of the local error surface. This procedure increases the learning rate, but only to the extent that the network can learn without large error increases Learning by Momentum It can be easily seen that if the input is to be learn nonlinear then the error surface has many local minima on it. The final value of the network's set of weights severely depends upon the initial state of the network. If the initial condition of the network falls into the domain of the local minima then the network may settle in the local minima instead of giving right solution. The momentum allows the network to ignore small features in the errors surface with momentum a network can slide through 62

21 such a minimum. Momentum can be added to the back propagation learning by making the weight changes equal to sum of fraction of last weight change and a new change suggested by the back propagation rule. Back propagation with momentum can be expressed mathematically by AW (i, j)= me AW (i,j)+(l-mc). d (i) p (j) S Learning By Adaptive Learning Method Learning rate is an important parameter used in the learning procedure. First, the initial network output and error are calculated. In addition, at each successive epoch new weights and biases are calculated using the current learning rate. As with momentum, if the new error exceeds the old error by more than a predefined ratio, typically 1.04, the new weights, biases, output and error are discarded. In addition, the learning rate is decreased typically multiplying by 0.7. If the new error is less than the old error, the learning rate is increased typically multiplying by This procedure increases the learning rate, but only to the extent that the network can learn without large error. 5.70UTPUT OF RESULTS FORM EEG DATA FILES OF PATIENT NUMBER 15 (EEG data is available from Ruby Hospital, Pune. ) Data file loaded of patient number 15 Display of original EEG signal Result after different operations are, Zero crossing Frequency in Hz Amplitude and Duration Amplitude at Samples Amplitude Min/Max Interchannel correlation and lag Spike detection 63

22 5.7.1 EEG Signal Data of Patient Number 15. CH-0 CHv CH* T-1 T-2 Fig: 5.12 original EEG signal for analysis Frequency Amplitude and Correlation Measurement Zero crossing frequency: Zero crossing frequency is the average frequency of the selected sample time for each selected channel Table 5.1 Time in seconds; Zero-crossing Frequency in Hz CHANNEL Zero crossing frequency in (Hz) CH CH CH CH CH CH CH CH

23 Amplitude and duration: The amplitude and duration tool is used to measure the absolute amplitude and duration a waveform. The amplitude consist of the sample number, raw time and the instaneous amplitude of selected samples for each selected channel Table 5.2 Time in seconds; Amplitude in uv. CHANNEL CH-0 CH-1 CH-2 CH-3 CH-4 CH-5 CH-6 CH-7 Amplitude at Sample 2495 Amplitude at Sample 4726 Difference in Amplitude The Min/Max Amplitude is used to measure the height of an event by finding the difference in amplitude at its min and max deflection Table 53 Min/Max and Diff over range. Amplitude in uv to sec. CHANNEL MIN. MAX. DIFF.IN AMPLITUDE AMPLITUDE AMPLITUDE ACCURACY CH % CH-1-32.S % CH % CH % CH % CH % CH % CH % 65

24 The correlation and lag is used to measure the interchannel correlation and lag time events between selected channels Table 5.4 Lag Time: Correlation: Time in msec Lags only valid for highly correlated (>=0.80) channels to sec. CH : : :032-15: : : : : :0.44 0: :0.47 0: : :034 7: : : :0.47 0: : :031-15: : : : : :0.44 0: :0.49 0: : : : : : :0.49 0: :035-54:032 46: : :034 15: : :035 0: :0.49 0: :0.30-7: : : :032-15:0.49 0: : :0.44 7: : :034-46:034 0: :0.43 0: : :038 10:037 8:037-7:036-19:037-2:036-13:038 66

25 5.7.3 CEG Signal Contains Spike of 256 Points of Epoch CH-0 0:1.00 msec H msec msec (±1000 msec) t-0:00:00;00.00 d-0:00:00:10.00 Window«2S6 Overlappcd-Y Wmdows/Epoch»l r> NUM Fig: 5.13 EEG signal contains spike of 256 points of epoch Prism is comprised of Voltage plot spectrum, offers superior analytical capabilities and allows exploring data dipole analysis [40] Correlation Graph for CH-0 and CH-! CH-0(0:1.00msec) CH msec msec Wmdow-256Overiapped»Y Wlndows/Epoch»l b num Fig 5.14 correlation graph for CH-0 and CH-1 67

26 5.7.5 Voltage Plot EEG Data Voltage plot reflects the average of selected samples. Voltage plot offers the spatial propagation of the averaged events [3^ Recorded [3 arc T] 80Hz[sOHzj^J j Ex Tape *] Fig Voltage plot EEG data file ANN is applied for amplitude, frequency, and co-relation measurement for the EEG data of different patients. Back propagation algorithm is selected to increase the learning rate. The performance of algorithms is faster to improve the analysis of EEG signal. Detection methods of k-complex in EEG waves are discussed in next chapter number 6. 68