The effects of audio stimuli on auditory-evoked potential in normal hearing Malay adults

Participant/subjects

This study was performed on 15 right-handed Malay male adult subjects (mean age = 23.5 year, standard deviation = 3.2 years). They were tested and certified fit by the Otorhinolaryngology (ENT) department in a Medical Centre for normal hearing threshold sensitivity through the routine pure tone audiometry measurement. All of the subjects exhibited normal audiological presentation in both ears (air conduction thresholds 15 dB hearing level (HL) from 125 to 4000 Hz bilaterally, 25 dB HL at 6000 and 8000 Hz, and pure-tone averages (average 20 dB HL from 500 to 4000Hz).

The experimental protocols were approved by the Human Ethics Committee of the University of Malay. The study was carried out on the approval of the Ethics (Medical Ethics Committee, University Malaya Medical Center. Ethics Committee/IRB Reference Number: 1045.22).

Each participant submitted written consent form before the experiments. A simple mini-mental state examination test was also conducted before the experiment to ensure that the subject’s mental abilities, memory capabilities, and attention and language proficiency met the required standards.^[15]

All participants showed 100% awareness, with healthy normal abilities and responses. However, the test had accounted for more than 15 subjects. Certain subjects’ recordings suffered from artifacts, noises, recording calibration, and device setting problems. The study selected the recording signals for subjects that had the most successful, free of artifacts, and noiseless signals. It turned out that only 15 subjects successfully met these criteria.

Stimuli

The study consisted of two disparate types of auditory stimuli; pure tone frequency burst (1 vs. 4 kHz) and speech consonant– vowel (CV) (/ba/vs./da/), presented at approximately 85–90 dB SPL. The tone stimulus was 200 ms in duration, generated by a software program in MATLAB R2013b from (mathwork.com), with (fall time = 10 ms, plateau time = 190 ms) represented at two different frequencies; 1000 and 4000 Hz tone stimuli.^[2]

The/ba/and/da/tokens were characterized by their contrasting voiced/voiceless articulatory features of speech, where/ba/has a higher formant frequency and onset frequency of the formant transitions compared to/da/.^[12,16] The speech stimulus was recorded at 44100 Hz sampling rate from the natural speech produced by a female Malay speaker. The CVs were edited into 200 ms in duration by removing the vibration of the vocal cords portion, the final part of the steady-state vowel, and windowing of the offset. The stimuli were presented with a pseudorandomized oddball sequence of 80% standard and 20% deviant presentations, with an interstimulus interval of 800 ± 500 ms, and delivered through Sennheiser HD 428 closed circumaural headphones to both ears. The oddball paradigm is an experimental design procedure, in which a sequences of a repetitive auditory stimuli are frequently interrupted by a deviant stimulus. Deviant stimuli are hidden in a rare occurrence issue among a series of more common standard stimuli. In this study, the pure tone stimulus had standard stimuli of (1 kHz) and deviant stimuli of (4 kHz). Furthermore, the CV stimulus had a standard stimulus of (da) and a deviant stimulus of (ba). The stimuli presented were calibrated at ear level using a KEMAR ear-and-cheek simulator (G.R.A.S. Sound and Vibration, 43AG) and a type 1 integrating sound level meter (Norsonic, nor140).^[17]

The tone and CV stimuli contrast were delivered separately and tested in two trials. Each trial consisted of 350 stimuli, that is, 70 deviant stimuli and 280 standard stimuli. Thus, there were 140 deviant stimuli and 560 standard stimuli presented over two trials. The order in which the stimuli were presented ensured that there were 3–5 standard stimuli between each deviant ones. There was no counterbalance for this study, that is, the (1000 Hz/da) stimulus was always the standard, while the (4000 Hz/ba) stimulus was always deviant.

ERP recording

Subjects were seated in a comfortable armchair inside a sound-proof chamber. They were instructed to minimize, and if possible eliminate, any eye blinking and muscle movements. The recording was done in various sessions at ~35 min each. To ensure the continuation of passive listening conditions, written short stories were presented throughout the experiment. The recording was done at 500 Hz sampling rate using the wireless Enobio EEG/ERP acquisition system.^[18]

Data were recorded from eight Ag/AgCl electrodes mounted on Neoprene EEG cap and located over the following scalp sites: Four electrodes were located on the midline of the head, Fz, Cz, Pz, FPZ, and other four electrodes were located on the left-hand side of the scalp, F7, C3, P7, and T5 (according to the modified International 10–20 System).^[19] Electroencephalography activity from each electrode was recorded with the common mode sense active electrode, while the driven right leg passive electrode was referred to the linked mastoid. The recording device Enobio EEG/ERP provided an online filter. The online filter consists of a bandpass filter, with passband (2–40 Hz) second order Butterworth FIR.

ERP waveform and component analysis

After ERP data collection, the responses evoked by the standard and deviant stimuli for both stimulus types (pure tone and CV) first went through pre-processing to correlate the baseline drift and filtered offline at 2–30 Hz using second - order Butterworth FIR bandpass filter. These evoked responses were averaged for each trial separately, and then these average evoked responses were averaged another time, with other trials in a session. Some recording sessions contained more than two trials, and some session resulted in bad trials (corrupted by artifacts and noises). The averaged trials were taken from successful runs free from artifacts, noises, and clearly evoked auditory ERP signals. This averaging process was done for each used electrode, separately. However, the standard average responses excluded responses to the stimulus occurring immediately after the deviant stimulus and vice versa for the deviant average response. However, there are many cleaning methods that could be used to denoise the raw signals.^[20] The empirical mode decomposition (EMD) technique was used to denoise the recorded EEG signals. All standard and deviant waveforms evoked responses were initially denoised by the EMD technique and inspected visually.^[21]

The criteria used to determine ERP response presence or absence were^[1] using visual inspection where the ERP is present if individual ERP peaks were larger than the level of the pre-stimulus baseline^[2] and using ERP analysis included baseline-to-peak amplitude and latency comparison with a typical standard ERP waveform as described by Näätänen R, McPherson et al., and Davies et al.,^[22-24] where N1 and N2 were defined as the most negative peaks occurring 80–150 ms and 180–250 ms after the onset of stimulus, respectively. P1, P2, and P3 were also defined as the most positive peaks between 55–80 ms, 145–180 ms, and 220–380 ms, respectively.

MMN was defined between 100 and 250 ms. In some trials, P1 and P2 were below the baseline, that is, a negative value, in which case the latency of the peak was measured and the amplitude recorded as missing. All measurements reported here are from responses recorded at the Cz electrode since it was at this electrode where the ERP was largest, while other electrodes were used for the purpose of comparison.^[25,26] Furthermore, MMN waveform was extracted using the difference waveform, where the deviant and standard stimuli were used (deviant minus standard).

Segmentation and feature extraction

The averaged cortical auditory-evoked potentials (CAEP) signals were segmented individually into time segments according to the CAEP latencies components. In which, for the P1 (latency window 20–100 ms), N1 (latency window 60–160 ms), P2 (latency window 140–240 ms), N2 (latency window 160–300 ms), and P3 (latency window 240–420 ms) ^[4,22,27] The latencies were obtained visually and using automated latency detection algorithms. This was done for each stimulus responses separately.

Non-linear feature extraction methods were used in this study such as root mean square Hilbert (RMSHilbert), Kolmogorov– Sinai entropy (KolmogEnt.), Sample Entropy (SampleEnt.), and Approximate Entropy (ApproxEnt.). This was due to the fact that brain neurons are controlled by nonlinear phenomena, such as the threshold and saturation processes. Therefore, its behavior can be regarded as non-linear, and non-linear dynamic analysis can be regarded as an integral approach in detecting mental tasks because it provides more information compared to that reported by traditional linear methods.^[28] Approximate Entropy (ApproxEnt.) evaluates the instability of variation in the signal, while Sample Entropy (SampleEnt.) measures the regularity of physiological signals. Furthermore, Kolmogorov– Sinai entropy (KolmogEnt.) evaluates the uncertainty of any signal with respect to time.

Classification of auditory ERP

Classification helps us to understand how a certain auditory brain response is related to other brain responses. Therefore, the classification of auditory response is crucial for patients suffering from hearing loss, as well as people with normal hearing.

The performance of the classifier is tested on different sets of instances. A cross-validation method was used to determine the trained and tested sets. The cross-validation process could be done through multiple approaches (i.e., K-fold cross- validation, Holdout validation, etc.). This study used K-fold cross-validation, with k = 4. This will make 75% of the data in the classification matrix used to develop an automated system to get the features that used to train the classifier, while 25% were used to test the classifier performances. The feature matrix generated from the successful feature extraction method was 240 × 5 elements for both types of stimuli. The matrix contains 4 features (RMSHilbert, KolmogEnt., SampleEnt., and ApproxEnt.), 4 stimuli type (1 kHz, 4 kHz, da and ba) for 15 subjects CAEP’s response, and for 5 time intervals, which were (P1, N1, P2, N2, and P3) intervals.

Therefore, for a classification matrix consists of 240 × 5 elements and using 4-fold cross-validation. The training matrix was 180 × 5 and the test matrix was 60 × 5, whereas the test matrix used to evaluate the classification accuracy; classification accuracy given a good indicator of the human brain responses to different auditory stimulus. However, to determine the brain’s ability to distinguish between the auditory evoked responses resulting from two auditory stimulus contrasts; pure tone frequency burst (1 kHz vs. 4 kHz) and speech consonant–vowel (CV) (/ba/versus/da/), and if there is a significant difference in the ERP to both stimulus.^[29] The classifiers used here to classify the brain responses (CAEP signals) into different auditory stimuli. In this study, a pure tone and CV stimuli were used as the auditory stimuli to evoke the brain ERP. A SVM with kernel function to learn and classify and K-nearest neighbor (KNN) classification methods were used to distinguish between the auditory-evoked responses to both stimulus contrasts. The SVM used a kernel trick to transform the data points into a higher dimensional space, and then separate them by a hyperplane at a maximal margin. The KNN determined a testing sample’s class by the majority class of the k nearest training samples.^[30,31] A feature extraction methods were applied for the EEG signals in the time domain, which were sample entropy, approximate entropy, Kolmogorov entropy, and RMS Hilbert transformation, these features were non-linear features.^[32]

In other hands, the classifier’s performance were determined using performance parameter (accuracy),^[33] defined as:

Statistical analysis

Statistical t-test was used to determine the significant difference in CAEP components due to both auditory stimuli, and to distinguish of the effective CAEP component related to each stimulus exclusively.

CAEP components

The all averaged CAEP responses for the tone stimulus (4 kHz) are shown in Figure 2. However, a sample of the average ERP waveforms for both contrast stimuli is shown in Figure 3.

Close

Figure 2:

The all averaged cortical auditory-evoked potentials responses for the tone stimulus (standard stimuli, Cz)

Export to PPT

Close

Figure 3:

A sample of the average event-related potential waveforms

Export to PPT

The mean and standard deviation of the peak amplitudes and latencies of the P1, N1, P2, N2, and P3 components for each group are shown in Table 1. However, the percentages of appearance for the ERP components in this work were as follows: P3 was present in 100%, N2 was present in 100%, P2 was present in 83%, N1 was present in 70.8%, and P1 in 73% of all average CAEP responses.

Table 1: Mean amplitudes in (µV) and latencies in (ms) with standard and deviations for the recorded ERP components with different stimulus (pure tone and CV stimulus) with normal hearing Malay subjects

In general, the results showed that all the average amplitudes of the CAEP components (P1, N1, P2, N2, and P3) were larger in standard stimuli compared to deviant stimuli for both stimulus types (pure tone and CV). Furthermore, the CAEP components P1, P2, N2, and P3 with response to CV stimulus showed larger latencies in standard stimuli compared to the deviant stimuli. The only exception was N1 wave, which exhibited the opposite pattern. While in the response to frequency stimuli, the CAEP components (P1, N1, P, and P3) showed larger latencies in responses to standard stimuli compared to responses to deviant stimuli. The only exception was N2 wave, which exhibited the opposite pattern. Furthermore, a t-test used for distinguishing the significantly effected CAEP component, which evoked by the auditory stimulus used on Malay subjects. Table 2 shows the t-test results for the amplitudes and latencies of the CAEP components in both stimulus type(s). From Table 2, we can infer that the P1 had a significant effect in amplitude due to CV stimuli, and there was no significant effected CAEP component elsewhere.

Table 2: The t-test results for the amplitudes and latencies of the CAEP components in both stimulus type

Classification

The training and testing were conducted in four times and the classification accuracy is averaged over the 4 trials. Table 3 listed the classification accuracies obtained from the two classifiers that used in this study. The non-linear feature extraction methods that used in this study resulted in a good classification accuracy.

Table 3: The performance parameter of classifier

However, the classification accuracy results show that the SVM classification analysis differentiated the groups of stimulus, pure tone frequency stimulus, and CV stimulus with high accuracy, revealing important interrelationship between the components. Furthermore, the kernel type for the support vector classifier was very efficient and separated the two classification groups by a good margin. Thus, it resulted in a good classification accuracy. The performance of the classifier was assessed by measuring its accuracy (accuracy = 84.58%).

Effects of frequency tone on ERP components

The tone frequency stimuli with a higher frequency (4 kHz) elicited a smaller amplitude response and smaller latencies relative to that of the lower frequency tone (1 kHz), which elicited a larger amplitude responses with bigger latencies. This finding agrees with previous studies. There are already some evidence from previous studies on MEG,^[41] PET,^[42] and FMRI,^[43] which reasoned that the area with neurons, located deep in the brain, responds better to high -frequency sounds, whereas the area with neurons, located superficially, responds better to low -frequency sounds. The only exception to this was the N2 wave, which showed different patterns with lower latency with the lower frequency stimulus. This study used a passive listening protocol, and the N2 component could identify in varieties of cognitive tasks including selective attention and stimulus change.^[39]

The results also showed unclear P1-N1-P2 complex in some subjects, where sometimes N1 and P1 were missing (the overall mean showed N1 was present in 70.8%, P1 in 73%, and P2 in 85% of all subjects ). This might be due to the fact that the P1-N1-P2 complex is sensitive to frequency stimuli^[44] and that the P2 amplitude dominates the response compared to P1 and N1 in adults.^[45]

P3 was elicited in all subjects, and it had a 100% appearance in both standard and deviant stimuli. Since the results were collected from the Cz electrode, P3 is regarded as P3b, which is usually recorded from the electrodes located over the centroparietal regions of the scalp (such as Cz and Pz sites).^[34,46] However, we preferred to use the term P3 in this article. An explanation of why P3 was the most detected component in this study is the fact that it is best elicited when the stimuli are delivered in oddball paradigm manner (for reviews see^[46] and^[47]).

Effects of CVS on ERP components

As documented in Table 1, the mean amplitude for all ERP components w as larger for responses to standard stimuli compared to the responses to deviant stimuli. One likely explanation for higher amplitude in responses to standard stimuli (/da/) is the spectral differences that exist in these CVs, where the formant frequencies and the onset frequencies of the formant transitions of /da /are lower than that of /ba/. Recently, the authors^[48] provided evidence using the FMRI technique on the area of the human cortex, located more superficially to the scalp, which responds better to low -frequency information, whereas the area located deeper in the brain responds better to high -frequency information. Therefore, the ERPs were recorded using surface electrodes and CVs with lower frequency energy that would mostly activate the more superficial regions of the scalp, thus producing larger amplitudes than CVs, dominated by higher frequency energies.

The mean latencies for all CAEP components were larger in the responses to standard stimulus compared to the responses to deviant stimulus for all used auditory stimuli. The present results agree with those reported in previous studies. First, we showed the latencies of CAEP responses to (85 –90) dB SPL / d a/stimuli being larger compared to responses to /b a/. Second, the amplitude of CAEP components was larger in /da /versus / ba /responses. This was due to the fact that brain neurons are responsive to sounds in certain specialized brain areas. This response depends on sound frequencies and amplitude contractions values.^[49] However, the only exception to this was the N1 wave, which showed different patterns with large latency in deviant CV stimuli as compared with standard CV stimuli. This is due to, N2 are an endogenous potential, and its appearance is not related to the physical property of stimuli, but it is related to human ’s reaction from the stimuli.^[50]

Classification accuracy

The accuracy of the results of the SVM classifier (which was 84.58%, Table 3) indicated that the human brain had a strong ability to distinguish auditory stimuli. However, the high accuracy of the SVM classifier gives us the opportunity to use different auditory stimuli in future to control certain applications using our minds.^[51] Furthermore, this classification protocol will support many systems that use the AEP signal classifications process in their application. For example, AEP signal classifications are used in BCI,^[8] brain hearing problems,^[9] and others.^[10] In fact, the limitation of sample size utilized in this study resulted in highly precise values for some of the utilized classifiers.