This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.
To classify Auditory Brainstem Response signals into three classes namely Normal, Early or Late using Neural Network classifiers.
During stimulation with sound, the Electro Encephalo Gram (EEG) undergoes typical changes that are related in time-locked manner to changes in the sound. (i.e.), they occur at a more or less fixed interval and with a similar waveform after the change in sound. These changes are called Auditory Evoked Potentials (AEPs). It is the response of the brain to any sound heard by the ear. Their voltage is generally smaller than the peak EEG voltage and can be recorded only after adding the changes to a large series of stimulus trials. This process, called averaging, makes use of the fact that the AEP has always nearly the same waveform, whereas the background EEG changes more randomly with respect to the stimulus change. As a result, the AEPs add up and the EEG slowly cancels, resulting in a steady growth of AEP amplitude with respect to the background noise.
Changes in the sound that reliably evoke AEPs are those in its amplitude and its frequency content. AEP tests are used to check the condition of the nerve pathways and help diagnose nervous system abnormalities, hearing loss, and assess neurological functions.
1.3 Auditory Brainstem Response
ABR is the first part of the AEP wave which occurs 1.5-15 ms post the presentation of a stimulus. There are five major waves in the ABR which are positive peaks generated from multiple axonal pathways in the auditory brainstem. The time between the onset of acoustic change and the occurrence of the peak voltage, called peak latency, of these AEPs ranges from 1ms to nearly 0.5 second. This means that even 0.5 second after a stimulus change, the changes in brain activity still are time-locked to that stimulus. Typically the brain's automatic responses to a sound occur within 50ms, but when active brain processing, such as detecting a wrong word in a sentence, is required, the responses occur later and later. Thus by comparing the measured peak latency values with the standard values, any abnormality in the signal can be detected.
1.4 Peak Detection and Classification
The absolute peaks of the ABR waves are detected by closely watching the zero crossings and the slope. After identification of the peaks and other features such as inter peak latencies, a neural network classifier is designed so as to automatically classify the responses as normal or abnormal.
2. LITERATURE REVIEW
Title of Paper
Analysis of parameters for the estimation of loudness from toneburst otoacoustic emissions
Michael Epstein, Ikaro Silva
Frequency specific stimuli of ABR is analyzed and the loudness-growth function is estimated through both OAE and ABR.
A new noise segmentation procedure is also
introduced that statistically partitions the background noise into dynamic discrete
segments of different noise power based on a series of F-tests
A preliminary attempt has been made at understanding the relationship between the performance of the evoked response loudness growth estimator and the estimated averaged evoked response residual noise levels. However, it remains to be seen if this measurement can be made across a wide enough range of frequencies.
An Analysis of Correlations among Abnormalities of ABR and Function of Brain
Mariko Fujikake , Satoki P.Ninomija,
Classification of ABR signals acquired from a group of brain damaged individuals into 5 classes was done by calculating eigen values of the signal.
The variables were taken into account after consulting with an expert. Too many parameters are considered and there is no information regarding the accuracy of the work.
Automated Analysis of the Auditory Brainstem
Andrew P. Bradley and Wayne J. Wilson
A peak detection algorithm for labeling the peaks and troughs of the ABR wave is done by finding out the first and second Gaussian derivatives.
All the significant peaks and troughs are detected and labeled with an accuracy of 96-98% with only 77% for wave IV.
Multi-Neural Networks Approaches for Biomedical Applications :
Classification of Brainstem Auditory Evoked Potentials
Anne-Sophie dujardin, Vkronique amarger, Kurosh madani
A neural based
biomedical diagnosis aide tool (BMDT)is designed. Quantization, Radial Basis Function and BackPropagation
neural networks are
The results can be different for different test sessions for the same patient, because they depend on
the relaxation of the person, the background, the test's
conditions, the signal-to-noise ratio. If the bad-classified vectors are introduced in the learning data-base, classification rates of both MNN
approaches decrease and become under 50%.
Nonlinear processing of auditory brainstem response
A. kworski, R.Tadeusiewicz, A. Paslawski
Automated identification method of wave V is presented, using the
multilayer perceptron type networks. Classification is based on cascade architecture of the neural
The neural network to the input of which a multidimensional
signal is being fed, must have more input elements than an equivalent network recognizing one dimensional signal. The increase in the number of connections leads to difficulties in realization of the network's learning process. When a limited
learning set is used and the network's memory
capacity is increased, deterioration of the network's properties with
respect to the generalization abilities happens.
Classification of the Auditory Brainstem Response (ABR) Using
Wavelet Analysis and Bayesian Network
Rui Zhang, Gerry McAllister, Bryan Scotney, Sally McClean, Glen Houston
In this study a method combining the
wavelet analysis and the Bayesian network is introduced to reduce the required number of
repetitions. The important
features of the ABR are extracted by thresholding and matching the wavelet coefficients.
These extracted features are then used as the variables to build up the Bayesian network for
classifying the ABR.
The numbers of averaged ABRs used at high stimulus levels indicate a considerable time
saving over pure coherent sub averaging, where no consideration is taken of apparent responses with minimum recordings. Whilst this could not be considered realistic at lower levels of stimulation it does have the potential to make a significant improvement in the overall efficiency of the technique.
Classification of auditory brainstem responses
by human experts and back propagation neural networks
Feed forward multilayered artificial neural networks
trained with the back propagation method are used to
classify auditory brainstem evoked responses (ABR). The
network response was moderate, approaching the lower
boundary of an optimum Bayesian classifier in several
cases. When human experts were forced to
classify the same samples, the performance rates were
The results suggest that while the knowledge of stimulus intensity is a very important factor in classifying ABRs, relative amplitude information
which contain indirect partial intensity clues is not useful. The difference in the performances of the networks trained on normalized and amplitude scaled data show
striking similarities to the changes in expert classifications on the same data. Furthermore, the features selected and
represented by the neural networks apparently closely
resemble those features of the signals used by the experts in making their classifications.
Optimal features for pedestal peak validation of ABR
The standard deviation of pedestal peak
signal and normalized amplitude of the pedestal peak are
considered as practical features to classify the pedestal peak.
Standard deviation of the filtered ABK signal is more
robust than the power ratio of filtered and unfiltered ABR signal to classify a pedestal peak signal
Use of instantaneous energy of ABR signals for fast detection
of wave V
Adeela Arooj, Mohd Rushaidin Muhamed, Sheikh Hussain Shaikh Salleh, Mohd Hafizi Omar
In this study, the instantaneous energy of ABR signal
had been introduced as a marker to identify the ABR waves rather than the conventional methods of FFT and wavelet transform.
A new method of ABR wave detection has been
designed. But the performance of this method needs to be tested further.
3. AUDITORY EVOKED POTENTIALS
3.1 Auditory System
The auditory system is the sensory system for the sense of hearing. A thorough understanding of the system is essential for carrying out trials with ABR.
3.1.1 Outer Ear
Sound waves are reflected and attenuated when they hit the outer ear (pinna), and these changes provide additional information that will help the brain determine the direction from which the sounds came. The sound waves enter the auditory canal, a deceptively simple tube. The ear canal amplifies sounds that are between 3 and 12 kHz. At the far end of the ear canal is the eardrum (or tympanic membrane), which marks the beginning of the middle ear.
3.1.2 Middle Ear
Sound waves traveling through the ear canal will hit the tympanic membrane, or eardrum. This wave information travels across the air-filled middle ear cavity via a series of delicate bones: the malleus , incus and stapes. These ossicles act as a lever and a teletype, converting the lower-pressure eardrum sound vibrations into higher-pressure sound vibrations at another, smaller membrane called the oval (or elliptical) window. The middle ear still contains the sound information in wave form; it is converted to nerve impulses in the cochlea.
3.1.3 Inner Ear
The inner ear consists of the cochlea and several non-auditory structures. Strikingly, one section, called the cochlear duct or scala media, contains an extracellular fluid similar in composition to endolymph, which is usually found inside of cells. The organ of Corti is located at this duct, and transforms mechanical waves to electric signals in neurons. The other two sections are known as the scala tympani and the scala vestibuli; these are located within the bony labyrinth, which is filled with fluid called perilymph. The chemical difference between the two fluids (endolymph & perilymph) is important for the function of the inner ear due to electrical potential differences between potassium and calcium ions.
3.1.4 Organ Of Corti
The organ of Corti forms a ribbon of sensory epithelium which runs lengthwise down the cochlea's entire scala media. Its hair cells transform the fluid waves into nerve signals. The journey of countless nerves begins with this first step; from here, further processing leads to a panoply of auditory reactions and sensations.
3.1.5 Hair Cell
Hair cells are columnar cells, each with a bundle of 100-200 specialized cilia, which are the mechano sensors for hearing. Lightly resting atop the longest cilia is the tectorial membrane, which moves back and forth with each cycle of sound, tilting the cilia, which is what elicits the hair cells' electrical responses. Hair cells show a graded response, instead of the spikes typical of other neurons. These graded potentials are not bound by the "all or none" properties of an action potential.
Afferent neurons innervate cochlear inner hair cells; at synapses where the neurotransmitter glutamate communicates signals from the hair cells to the dendrites of the primary auditory neurons. There are far fewer inner hair cells in the cochlea than afferent nerve fibers. The neural dendrites belong to neurons of the auditory nerve, which in turn joins the vestibular nerve to form the vestibulocochlear nerve, or cranial nerve number VIII. Efferent projections from the brain to the cochlea also play a role in the perception of sound.
Fig. 3.1 Structure of Ear
3.1.7 Central Auditory System
This sound information, now re-encoded, travels down the vestibulocochlear nerve, through intermediate stations such as the cochlear nuclei and superior olivary complex of the brainstem and the inferior colliculus of the midbrain, being further processed at each waypoint. The information eventually reaches the thalamus, and from there it is relayed to the cortex. In the human brain, the primary auditory cortex is located in the temporal lobe.
Fig. 3.2 Auditory Pathway
3.2 Evoked Potentials
EEG is the recording of the electrical activity of the brain that is produced by the firing of neurons. Evoked potentials are the derivatives of EEG recording wherein the activity is time locked to a stimulus which can be either a visual, auditory or somato-sensory stimulus.
3.2.1 Auditory Evoked Potential
The human brain consists of 1010 nerve cells (approx.) that are active continuously. (i.e.), they continuously undergo changes in the internal voltage, regardless of the state of the brain. The voltage changes in neocortical nerve cells can be easily recorded with the use of scalp electrodes and can be visualized on an oscilloscope or computer screen after suitable amplification. The combined electrical activity from a large number of cortical neurons recorded from the scalp is called Electro Encephalogram (EEG). The EEG changes as a function of the brain state and what is recorded also depends on the location of the electrodes.
Recording electrodes placed over the temporal lobes, (i.e.) between the ears and halfway up to the midline of the head, are in the best position to record changes specific to auditory cortex. However, electrodes near the vertex of the head are commonly used, and they pick up activity from both the hemispheres, including both temporal lobe activity and that from attention centers in the frontal cortex. Activity from deeper brain sources, such as those in the brainstem, also is recordable, although with much lesser amplitude.
3.2.2 AEPS and Brain Structures
The study of the auditory system on the basis of evoked potentials is concerned with the effects of a sound on the neural activity along the auditory pathway. These effects range from the initial changes in the membrane potential of the hair cells in the cochlea to the detection of a semantic mismatch in a series of words. In fact, with AEPs one can trace the neural activity as it makes its way to the auditory cortex and then cycles back and forth between various auditory cortical areas and between those areas and the thalamus to produce the long-latency components. EP generated in cochlea goes through cochlear nerve, cochlear nucleus, superior olivary complex, lateral lemniscus, inferior colliculus, medial geniculate body and then to the cortex.
Fig. 3.3 Illustration of Auditory Pathway
The auditory cortex is a hierarchal structure in which one can distinguish three subdivisions. In humans, one typically uses primary and secondary cortex to distinguish core from other areas. The core areas each receive input from specific auditory areas in the thalamus. This excitatory input typically arrives on the dendrites when the axons of the thalamic cells pass through layer iii or iv. Neural activity from auditory nerve fibers and fiber tracts in the brainstem gives rise to small far-field potentials that can be recorded from the scalp. These far-field potentials represent stationary peaks resulting from a slowly moving source.
3.2.3 Components of AEP Wave
The AEP wave form consists of 3 parts:
1. Auditory Brainstem Response (ABR)
ABR is the first part of the AEP wave as it originates in the VIII cranial nerve and brainstem auditory structures. It occurs 1.5-15ms post stimuli and includes waves I-V.
2. Middle Latency Response (MLR)
It occurs 25-50 ms post stimulus. It originates in upper brainstem or auditory cortex and includes waves Na and Pa.
3. Late Response (LR)
It occurs 50-200ms post stimuli and includes the P1-N1-P2 sequence.
Fig. 3.4 Components of AEP Wave
3.2.4 The ABR
ABR is evoked when a stimulus click is applied to a subject's ear to determine hearing acuity and integrity of the auditory pathways. If the stimulus is perceived, a response changes their EEG within 10ms from stimulus onset. The amplitude of the ABR signal is 1microvolt-5microvolts and is hidden behind the background EEG and noise (50microvolts approximately).
The components of the ABR are swamped by the electrical activity of the brain and the determination of a response can be difficult particularly at low levels of auditory stimulation, as hearing threshold is reached. The ABR waveform is extracted by coherent averaging, which exploits the deterministic nature of the signal, to enhance the waveform while suppressing the uncorrelated EEG, extraneous noise and artifact. It is necessary to average approximately 1000-2000 trials before the noise is sufficiently suppressed, with signal to noise ratio (SNR) enhanced proportional to the square root of the number of trials.
188.8.131.52 Waveform Components
i) Wave I: The ABR wave I response is the far-field representation of the compound auditory nerve action potential in the distal portion of cranial nerve (CN) VIII. The response is believed to originate from afferent activity of the CN VIII fibers (first-order neurons) as they leave the cochlea and enter the internal auditory canal.
ii) Wave II: The ABR wave II is generated by the proximal VIII nerve as it enters the brain stem.
iii) Wave III: The ABR wave III arises from second-order neuron activity (beyond CN VIII) in or near the cochlear nucleus. Literature suggests wave III is generated in the caudal portion of the auditory pons. The cochlear nucleus contains approximately 100,000 neurons, most of which are innervated by eighth nerve fibers.
iv) Wave IV: The ABR wave IV, which often shares the same peak with wave V, is thought to arise from pontine third-order neurons mostly located in the superior olivary complex, but additional contributions may come from the cochlear nucleus and nucleus of lateral lemniscus.
v) Wave V: Generation of wave V likely reflects activity of multiple anatomic auditory structures. The ABR wave V is the component analyzed most often in clinical applications of the ABR. Although some debate exists regarding the precise generation of wave V, it is believed to originate from the vicinity of the inferior colliculus. The second-order neuron activity may additionally contribute in some way to wave V. The inferior colliculus is a complex structure, with more than 99% of the axons from lower auditory brainstem regions going through the lateral lemniscus to the inferior colliculus.
vi) Wave VI and VII: Thalamic (medial geniculate body) origin is suggested for generation of waves VI and VII, but the actual site of generation is uncertain.
Fig. 3.5 Components of ABR Wave
184.108.40.206 Applications of ABR
The ABR has proven useful in the detection of vascular lesions of the brainstem, including hemorrhage from a ruptured blood vessel or interruption of blood flow because of occlusion of a blood vessel. The ABR has been used in the differential diagnosis of comatose patients. The ABR will typically not be affected if coma is due to metabolic or toxic causes but will show changes if coma is due to a structural lesion. ABR testing can help to determine the extent of brain damage in head trauma victims or infants suffering from lack of oxygen because of birth complications and is used in the determination of brain death.
Fig. 3.6 ABR Testing in an Infant
In otoneurological applications, the ABR is useful in assessing the status of the auditory nerve and brainstem pathways because damage to these areas can alter the ABR in characteristic ways. The location of the lesion will affect the ear in which ABR abnormalities are manifested. In general, a lesion of the auditory nerve will affect the ABR generated by the stimulation of the ear ipsilateral to the lesion.In cases of large tumors located at the cerebello-pontine angle, contralateral effects may also be seen due to compression of the brain stem. Lesions of the brain stem may cause ipsilateral, contralateral, or bilateral abnormalities of the ABR. In using ABR to assess the integrity of the central auditory pathways, click stimuli are generally used because they elicit the clearest waveform.
ABR testing with click stimuli can predict auditory sensitivity within the 1-4 KHz range to within 5-20 dB. A major limitation of click stimuli is that they are not frequency- specific and hence they cannot provide information about the entire audiogram. In addition, because click threshold gives an estimate of hearing within the mid-frequency to high-frequency range, it is possible to have a normal ABR and still have significant hearing loss. Specifically, click ABR will not be sensitive to hearing loss below 1 kHz or above 4 kHz. Hence tone burst stimuli are preferred for recording the ABR signal .
Absolute latency is the difference in time between presentation of a stimulus and occurrence of the peak wave. Inter peak latency is the time difference between two wave components. The peak latencies were first discovered by Jewett and hence they are also commonly known as Jewette peaks. Latencies and inter peak latencies are important features for any kind of ABR analysis. The standard values of these latencies are given in the table below.
LATENCY/INTER PEAK LATENCY
Table 3.1 Standard Latencies
Initially, EEG recording was done on healthy subjects using a standard EEG acquisition system. Scalp electrodes were placed according to the international 10-20 electrode placement system. Pure tone stimulus from an audiometer was given via headphones to the subject at a fixed intensity of 60 dB. Frequency of the stimulus ranged from 1000-8000 Hz and many such trials were carried out. The acquired signals were preprocessed and sufficient noise artifacts were removed. Extraction of AEP from the EEG signal using adaptive filtering and wavelet decomposition was attempted. Due to the difficulty in filtering out the EOG spikes and time constraint, the signals were taken from the MIT database.
Fig. 4.1 EEG Acquisition
4.1 Data Description
The ABR signals were downloaded from the MIT Physionet Database. The experiment has been conducted by affixing three electrodes on the subject and ABR was recorded in a sound proof room. The location of the placement of electrode was cleaned with alcohol and the non-inverting electrode was positioned on the forehead, the inverting electrode was positioned on the ipsilateral mastoid (behind the ear), and the ground electrode positioned on the contralateral mastoid.
The acquired signal was amplified using a GRASS QP511 Quad AC Amplifier with an amplification factor of 50000 and a band pass filter with cut-off frequency of 30-3000 Hz was applied. The signal was sampled at 48000 Hz. The effective resolution of the recording was 24-bits. Stimuli were presented at a presentation rate of about 24 Hz (2002 samples/trial). Eight subjects were used for the recording and the stimulus was given from the threshold of hearing of the subject until 100 dB. The frequency of the stimulus was 1 KHz or 4 KHz.
Fig. 4.2 Acquisition of ABR
The acquired signal was smoothened using a smoothening filter and the frequency content of the signal was analyzed by plotting the absolute value of the Fast Fourier Transform (FFT). Since band pass filtering was already performed, no significant high frequencies and noise were found out.
The stimuli given were in the form of a tone burst. Therefore many events occur simultaneously. Hence segmentation is done and the appropriate segment is then considered for feature extraction.
Fig. 4.3 Spectrum of ABR
Fig. 4.4 Smoothening
4.3 Peak Detection
The segment chosen is plotted in the sample scale and a threshold is fixed. To detect the peaks, the envelope of the signal above the threshold is plotted. The peaks appear to be distorted. Therefore the segment is smoothened. Now the peaks alone are plotted in the sample scale. To detect the time at which the peaks occur, amplitude values that occur in adjacent samples are compared to determine if there is any zero crossing. The direction of the zero crossing will indicate either a peak or a trough. Peaks are of primary concern. If the compared values keep on increasing, the signal is an ascending slope. When the values begin to decrease, the slope becomes a descending one and a peak is detected. Similarly all the other peaks are detected whenever an ascending slope turns into a descending slope. After the peaks are detected the difference between waves I-III, III-V and I-V are calculated as inter peak latencies and the results are tabulated.
Fig. 4.5 Envelope Detection
4.4 Feature Extraction
Four statistical features and six morphological features were taken into account for the purpose of classification. The statistical features that were calculated are mean, variance, power and energy. The morphological features that were calculated are three absolute latencies and three inter peak latencies. The results are tabulated in an excel sheet. The last column of the sheet is filled with the target values. These features are now ready to be given as input to a neural network.
Fig. 4.6 Statistical Features
Fig. 4.7 Peak Position Indication
4.5.1 General Algorithm of MLP
The Feed Forward Back propagation network consists of N layers- the input layer, two or more hidden layers and an output layer. The first layer gets the weights from the input. Subsequently, each layer gets the weights from the previous layer and every layer also has a Bias associated with it. The final layer's output is the network output. With each iteration the weights of each neuron are updated with respect to a particular learning function. This is called the Network's adaptation. A specific training function trains the neural network and the performance of the network can be measured by a performance function.
Fig 4.8 General Structure of MLP
A Feed forward MLP classifier is built that can classify the ABR response into normal, early and late from features- both statistical and morphological. The ten features serve as the inputs to a neural network and the type of the ABR responses such as normal, early and late are the targets. To achieve this, training of the MLP network is done by presenting previously recorded inputs and then tuning it to produce the desired target outputs. The training algorithm used is the Levenberg-Marqaurdt algorithm.
The samples are divided into training and test sets. The training set is used to teach the network. The test set provides a measure of network accuracy. The network response is then compared with the desired target response. This creates a classification matrix which gives an idea of the classifier's performance. In the ideal classification matrix the sum of the diagonal should be equal to the number of samples. 352 records are taken under training set and 100 records under testing set. The number of neurons in the hidden layer and the number of training iterations are varied and for each case the classifier accuracy is noted.
Fig. 4.9 Network Training Tool
4.6 Results and Discussion
Fig. 4.10 Classification Output
Fig. 4.11 Performance Plot
Fig. 4.12 Training State Plot
After training, the neural network achieves the best validation at epoch 9 and the validation curve reaches the best state within 15 iterations. This shows that the efficiency of the network is good. The percentage of correct classification is also high (91%) which shows the classifier's efficiency.
4.7 Performance Evaluation
Statistical measures of the performance of any classifier are essential to evaluate its efficiency. The three important statistical parameters are calculated for this neural network classifier as follows.
i) Sensitivity is the measure of proportion of actual positives which are identified correctly. It is calculated using the formula
Sensitivity= (True Positives) / (True Positives + False Negatives)
ii) Specificity is the measure of proportion of actual negatives which are correctly identified. It is calculated as follows.
Specificity= (True Negatives) / (True Negatives+ False Positives)
iii) Accuracy indicates proximity of measurement results to the true value. It is calculated as follows.
Accuracy = (True Positives + False Negatives) / (Total Number of Samples)
TABLE 4.1 Performance Evaluations
The statistical and morphological features of ABR waveforms such as mean, variance, energy, power latencies and inter peak latencies of the peaks were extracted using the peak detection algorithm. The features were given to an MLP classifier with 20 hidden neurons. The classifier was trained with 352 records and tested on 100 records. Hence from the above results, it can be concluded that an MLP based ANN Classifier shows a superior performance of classification based on the features fed to the network. Hence automatic classification of ABR waveforms into normal, early or late classes can be done and in case of any abnormality, suggestive pathologies can be indicated.