In this research, we mainly verified two aspects:
Experiment 1. Differentiating between the patients with tinnitus and healthy controls
Experiment 2. Verification of the distress level of patients with tinnitus
The institutional review board of Keio University Hospital and the bioethics committee of Keio University Science and Technology approved this study (UMIN000013657). The research was
performed in accordance with clinical research guidelines in Japan, and informed consent was obtained from all patients.
Experiment 1: Classification of patients with tinnitus and healthy controls
To verify whether there is a difference between EEG signals of patients with tinnitus and healthy controls, the EEG features of both groups were obtained and classified using only the detected features. First, filtering of the EEG data was performed as a preprocessing a step and normalization was performed after fast Fourier transformation. We then compared the amplitude spectra of the frequencies in the patients with tinnitus and healthy controls to investigate whether there was a significant difference. To demonstrate the effectiveness of the frequency features, we calculated the recognition accuracies using support vector machine (SVM), which focused on the frequency difference between those with and without tinnitus.
2.1) Participants
Overall, 31 patients with tinnitus (mean ± SD age, 63.8 ± 8.6 years; 19 males and 12 females) and 29 healthy controls (mean ± SD age, 68.2 ± 10.4 years; 15 males and 14 females), participated in this study (Table 1). No significant differences in age were found between the two groups (p > 0.05). In addition, no significant difference in the hearing threshold was found between the two groups (p > 0.05; Figure 1).
Table 1
Distress of the patients using THI
THI score
|
Distress
|
The number of patients
|
0-16
|
Nothing
|
3
|
18-36
|
Small
|
9
|
38-56
|
Moderate
|
9
|
58-100
|
High
|
10
|
2.2) Tinnitus handicap inventory
After the EEG measurement, the degree of distress experienced in everyday life by each patient with tinnitus was evaluated using tinnitus handicap inventory (THI)22,23. THI is scored on a scale of 0 to 100. Table 1 shows the relationship between THI score and severity of tinnitus22. From the experiments, we obtained EEG data, and information on the degree of tinnitus distress was obtained from responses to the questionnaire (Table 1).
2.3) EEG measurement
The participant sat on a chair wearing the EEG device (Figure 2). We then recorded the resting state EEG signals for 30 seconds with the participant’s eyes closed. EEG measurements were conducted twice, including a 30-second break. The experimental flow is shown in Figure 3. The measurement point was only Fp1 as per the international 10-20 system. The sampling frequency was 1024 Hz.
2.4) Data Analyses
2.4.1) Preprocessing
Using a low-pass filter with a cutoff frequency of 30 Hz, we obtained EEG data without noise.
2.4.2) Feature extraction of data
Frequency analysis of the preprocessed EEG data was performed using fast Fourier transformation, the amplitude spectrum data were acquired, and the components of 4–22 Hz were extracted. When the frequency analysis was performed, the time window was set to 1 s, and a Hamming window was applied to the EEG data. The frequency resolution of the obtained amplitude spectrum was 1 Hz. In this study, 19-dimensional amplitude spectrum data in the 4- to 22-Hz interval were used. As the length of the time window and amount of shift was 1 s, the number of data obtained from one EEG measurement was 30. As we measured each subject twice, we obtained 60 data per person. The 60-amplitude spectrum data of each subject were averaged for each frequency. As the amplitude of EEG differed in each person, normalization was performed to eliminate individual differences. As a method of normalization, the sum of the amplitude spectra of 4–22 Hz was set to 1. This method is often used to compare the ratio of the amplitude of each frequency. Using the obtained amplitude spectrum data, the mean and standard deviation of the 60 data of all the patients with tinnitus and all healthy controls were calculated for each EEG frequency.
2.4.3) Statistical analysis
To investigate whether a frequency component with a difference in EEG exists between the patients with tinnitus and healthy controls, the statistical significance (t test) was examined for each frequency in the amplitude spectrum data. The statistical significance was set at <0.05 (two-tailed). In this case, when the significance level is 5%, and the p value is >0.025, it indicates no statistically significant difference between the two groups. If the p value is <0.025, a significant difference exists.
2.4.4) Classification
Classification is performed using features that show the possibility of a difference between the patients with tinnitus and healthy controls. In this study, we used the RBF kernel as a kernel function for classification using SVM. The optimum setting of the parameters γ and C often poses a problem. To set the optimal parameters, we adopted the method recommended by Chih-Wei Hsu et al.24. Classification was repeated while changing the possible values of parameter C to・・・ and changing the possible value of parameter γ to, , ・・・.
This allows for the setting of the parameters to roughly increase the classes. Next, the range of the parameter searched was limited only to that parameter with the highest identification rate. Within the limited range, classification was repeated while changing the numerical value of each parameter. Using this procedure, the optimum parameter was set.
The recognition accuracy was calculated using the extracted features. If we obtain a high recognition accuracy, we could extract the difference between the patients with tinnitus and healthy controls. If a difference existed between the EEG of the two groups, the effect of tinnitus would appear in the EEG data of the prefrontal cortex. In this research, the leave-one-subject-out cross-validation (LOSOCV) method was used to calculate the recognition accuracy. The data of one person was used as the evaluation data and those of the remaining participants were set as the learning data. The recognition accuracy was calculated for all the combinations and was evaluated by setting the obtained recognition accuracy average value as the final recognition rate.
Experiment 2. Validation of distress level
To verify whether tinnitus severity can be identified on the basis of brain waves, we investigated the relationship between the EEG data and THI scores of the patients with tinnitus to classify their levels of distress. For this purpose, we performed the same procedures as in Experiment 1.
2.5)Data analysis
2.5.1) Statistical analysis
To investigate the difference in EEG data of the patients with tinnitus due to the difference in pain level, a significant difference test was performed using the amplitude spectrum data. The null hypothesis was “there is no difference between the data of group A and group B.” Groups A and B were defined in three ways as shown in Table 2.
Table 2
p values at each frequency of the EEG
Frequencies(EEG) [Hz]
|
4
|
5
|
6
|
7
|
8
|
\(p\)value
|
0.0732
|
0.0631
|
0.320
|
0.575
|
0.636
|
Frequencies(EEG) [Hz]
|
9
|
10
|
11
|
12
|
13
|
\(p\)value
|
0.403
|
0.751
|
0.131
|
0.0890
|
0.00101
|
Frequencies(EEG) [Hz]
|
14
|
15
|
16
|
17
|
18
|
\(p\)value
|
0.0470
|
0.719
|
0.853
|
0.681
|
0.390
|
Frequencies(EEG) [Hz]
|
19
|
20
|
21
|
22
|
|
\(p\)value
|
0.511
|
0.752
|
0.212
|
0.211
|
|
To investigate the difference in EEG due to distress, validation was performed except for the participants with a THI score of 0.
2.5.2) Single regression analysis
Single regression analysis was used to verify whether the distress caused by tinnitus can be estimated using EEG. A single regression analysis was applied to the different frequencies to construct a regression equation for estimating tinnitus distress. To verify how the amplitude spectrum of the EEG changes relate to changes in the distress level, a regression equation was obtained. At this time, the target variable was the amplitude spectrum of the EEG, and the explanatory variable was the THI score.
2.5.3) Classification
Distress in the patients with tinnitus was identified using EEG features correlated to distress. The participants for comparison were the three types of patients with tinnitus shown in Table 2. We set the parameters and calculated the recognition rate in the same manner as in Experiment 1.