CORRELATION BETWEEN THE HEARING LOSS CLASSIFICATIONS AND SPEECH

(1) Federal University of Minas Gerais – UFMG, Belo Horizonte, MG, Brazil. (2) Federal University of Minas Gerais – UFMG, Belo Horizonte, MG, Brazil. (3) Speech-language pathology and Audiology, Federal University of Minas Gerais – UFMG, Belo Horizonte, MG, Brazil. (4) Speech-language pathology and Audiology, Federal University of Minas Gerais – UFMG, Belo Horizonte, MG, Brazil. Conflict of interest: non-existent Those afflicted by hearing impairments tend to avoid contact, develop low self-esteem, have few friends and limited participation in social relations3. Ski slope sensorineural hearing loss is the most common type to be found in clinical audiology practice; it includes presbycusis and is frequently related to difficulties in speech intelligibility4-6. The frequencies 500 Hz, 1000 Hz and 2000 Hz are considered to be the most relevant for speech. Vowels and consonants, the building blocks of speech, have different spectral characteristics. Vowels are naturally more intense and carry acoustic energy at low frequencies (400–500 Hz), which are favored by the range of human audibility7. Consonants, in turn, are sounds having spectral energy at high frequencies, above 2000 Hz, albeit 20–35 dB weaker than vowels8. However, speech intelligibility depends on consonant sounds, which  INTRODUCTION

correlation with the speech recognition threshold (SRT) and with the speech discrimination score (SDS).

METHODS
The present study was approved by the Research Ethics Committee of the Federal University of Minas Gerais (UFMG) under protocol no.155 09.An informed consent form was prepared explaining the theme of the study, its aims and the importance of analyzing the audiometric tests of the individuals selected for the study.
This comparative study based on the results of audiometric tests was conducted with an elderly population who received care at the Instituto Jenny de Andrade Faria, an annex to the UFMG Hospital das Clínicas.
The study employed a convenience sample of all the elderly individuals with a complaint of hearing loss who underwent audiometric evaluation with measurements of pure-tone hearing thresholds for air and bone conduction and speech audiometry consisting of the SRT and SDS tests between April 2011-April 2012.Some exclusion criteria were established: individuals with mixed or conductive hearing loss in at least one ear, or incomplete/ inconclusive tests, e.g, when pure-tone thresholds over the frequencies of 500 Hz to 4000 Hz were not obtained or speech audiometry was not performed.
The present study examined the pure-tone hearing thresholds for air conduction obtained through the analysis of pure-tone audiometry based on the following averages:   The data were treated statistically using the R software.Distribution of frequencies was performed for the categorical variable (sex).For the continuous variables (age, frequencies, SRT and SDS), measures of central tendency (mean and median) contribute 60 % of it, while vowels contribute only 40% 9 .Due to the spectral characteristics of these sounds and the range of human audibility, one can understand why individuals with hearing loss at high frequencies have impaired speech recognition.
In audiology, speech recognition ability is measured by the tests Speech Reception Threshold (SRT) and Speech Discrimination Score (SDS).The SRT corresponds to the softest sound intensity level at which an individual can recognize 50% of the common words given.Typically, SRT and SDS values are consistent with the average of the hearing thresholds obtained for the speech-related frequencies.The SDS evaluates speech discrimination using a list of monosyllables and bisyllables 40 dB above the SRT thresholds 10 .
The classification of hearing losses is a widely discussed topic in speech-language pathology and audiology.In Brazil, hearing losses have been predominantly classified on the basis of the frequencies of speech relative to the tritone average of the air-conduction thresholds for 500 Hz, 1000 Hz and 2000 Hz 11,12 .However, a number of authors have advocated the use of hearing loss classifications by frequency ranges in order to include the high frequencies in composing these averages, especially for elderly individuals 13,14 .
The Bureau Internacional d' Audiophonologie (BIAP, 2005) recommended that the classification of hearing losses should take into account the averages of the pure-tone thresholds for air conduction at 500 Hz, 1000 Hz, 2000 Hz and 4000 Hz in order to encompass the high frequencies, as these are the most compromised in the pathologies of hearing 15 .
Because the frequency ranges of consonants above 2000 Hz are the chief constituents of speech intelligibility, Russo proposed, in 2009, that the degree of hearing loss should be classified based on the average audibility threshold obtained over the frequencies of 500 Hz to 4000 Hz 13 .
The classifications of hearing losses relying on the tritone average as suggested by Lloyd & Kaplan  (1978)  12 and Davis & Siverman (1970) 11 were found to be adequate to classify the hearing losses with a flat audiometric configuration.However, in ski slope hearing losses, such as presbycusis, this classification is not always consistent with the patients' complaints, as these refer to difficulties in speech recognition, mostly.
Given that the major auditory function concerns verbal communication and speech recognition abilities, it is paramount that the speech tests be considered in the classification of hearing losses, thus expressing the actual hearing impairment of the patients.The aim of the present study was to assess which pure-tone averages have a stronger coefficient of correlation, i.e., the greater the correlation, the lower the MSE (lower prediction error); consequently, the better the variable as a predictor.
To ascertain whether the correlations obtained were statistically different, comparison of correlations was the method adopted for hypothesis testing.The correlation values were compared both for the SRT and the SDS, and the results were obtained using p-values, with statistical significance defined by p<0.05.

RESULTS
The mean values obtained in the audiologic evaluation of the 482 ears can be found in Figure 1.
and variability (standard deviation, minimum and maximum) were used.
The four averages and the individual frequencies 500 Hz, 1000 Hz, 2000 Hz, 3000 Hz and 4000 Hz were correlated with the SRT and SDS values obtained in the audiometric tests in order to assess which pure-tone average or individual frequency best represents speech recognition ability.To that end, Spearman's coefficient of correlation was used, with a confidence interval of 95%.To determine the strength of the correlation, the following classification scale was used 16 : 0-0.2: very poor correlation; 0.21-0.4:poor correlation; 0.41-0.6:moderate correlation; 0.61-0.8:good correlation, and 0.8 -1.0: very good correlation.
In addition, the mean squared error (MSE) was calculated.The MSE is inversely related to the Table 1 shows the descriptive analysis of the threshold averages by frequency, the four calculated pure-tone averages, the SRT in dBHL, and the SDS in percent values.

DISCUSSION
The aim of the present study was to identify which pure-tone average has the greatest correlation with speech recognition.To that end, the pure-tone thresholds for air conduction were classified using four different averages: Average 1 (500 Hz, 1000 Hz and 2000 Hz), average 2 (500 Hz, 1000 Hz, 2000 Hz and 4000 Hz), average 3 (500 Hz, 1000 Hz, 2000 Hz and 3000 Hz) and average 4 (500 Hz, 1000 Hz, 2000 Hz, 3000 Hz and 4000 Hz).These averages were correlated with the tests of speech recognition, SRT and SDS, obtained in the audiologic evaluation.
The curve depicted in Fig. 1, which shows the average pure-tone thresholds of all the study patients by frequency, is similar to a typical audiogram of elderly individuals, as expected.In this population, a ski slope configuration is commonly found, with a high-frequency hearing loss and preserved low-pitch sensitivity 14,[17][18][19][20] .Presbyacusis is defined as a bilateral sensorineural hearing loss that is more pronounced for high-pitched sounds due to the degenerative and physiologic changes in the auditory system that arise with aging 17,21 .In elderly individuals with presbyacusis, the apical coil of the cochlea, responsible for the detection of low-pitched sounds, is more preserved compared with the basal coil, which detects the high-pitched sounds 17 .
Regarding speech discrimination, Table 1 shows that the mean for the SDS was 75.99%, which was expected for elderly patients [22][23][24][25][26] .Studies suggest that the difficulty in speech comprehension, demonstrated by the lower SDS, occurs as a result of the structural and physiologic changes in the auditory system with aging 22,23 .
Table 2 shows the correlation values for the individual frequencies 500 Hz, 1000 Hz, 2000 Hz, 3000 Hz and 4000 Hz, Average 1, Average 2, Average 3 and Average 4 with the SRT and SDS.Although all the averages showed very good correlation, it was found that Average 1 had greater correlation with the SRT (0.934).For that reason, Average 1 was considered to be the most adequate estimator of the threshold of speech recognition 16 .This fact was corroborated by the mean squared error found in the correlation of Average 1 with the SRT (52.2).This lower mean demonstrated the lower are more frequent in the patient's vocabulary, which increases redundancy and the chances of correct guesses.This makes guessing with the SRT test easier than with the SDS.
A recent study compared the audiologic performance of elderly individuals based on the classifications of Davis and Silverman, 1970 11 and on the Recommendation 02/1 of the Bureau Internacional d'Audiophonologie (BIAP) 15 .The authors reviewed the records of 140 elderly individuals focusing on the clinical history and the pure-tone audiometry results.Predominance of mild to moderate sensorineural hearing loss was found, with slight differences regarding its prevalence depending on the hearing loss classification adopted.Based on Davis and Silverman, 99 cases of ears with normal thresholds were found, while only 66 cases were identified according to the BIAP recommendation.The study concluded that the classifications achieved similar results; however, the BIAP recommendation was found to be more sensitive in detecting hearing losses in the elderly 27 .A study conducted in Finland with 5400 volunteers aged 55-75 years demonstrated a remarkable difference between the hearing loss classification by the World Health Organization (WHO) criteria compared with the European Union criteria.In that study, the percentage of individuals with normal hearing was greater when the WHO classification was used 28 .
The results of the analyses enabled us to conclude that speech discrimination as measured by the SDS is influenced by the frequencies 3000 Hz and 4000 Hz.For that reason, the importance of including these frequencies in the pure-tone average used for classifying the hearing loss should be emphasized, since one of the objectives of this classification is to reflect a person's hearing abilities and difficulties especially with regard to communication.

CONCLUSION
For the elderly population with ski slope sensorineural hearing loss, the SRT is more strongly correlated with the average of frequencies 500 Hz, 1000 Hz and 2000 Hz, while the SDS shows greater correlation with the average that includes frequencies 3000 Hz and 4000 Hz. prediction error of Average 1 relative to the SRT.In other words, Average 1 is the best predictor of SRT values.Additionally, Fig. 2 shows that the correlation of SRT and Average 1 differs statistically from the correlations between SRT and the other averages.This proves that, even with such narrow confidence intervals (Table 2), the correlation of Average 1 with the SRT can be regarded as the strongest.Thus, this result indicates that the averages of 500 Hz, 1000 Hz and 2000 Hz used to compose Average 1 are key frequencies to estimating the SRT.This fact can be confirmed by the analysis of the individual frequencies in Table 2.
The correlation values for the SDS are negative, since they are inversely proportional: as averages increase, SDS percentages decline.All the averages had good correlation with the SDS, with statistically superior correlation values for Averages 2, 3 and 4 (Table 2 and Fig. 2).As Fig. 2 illustrates, the correlation between Average 1 and the SDS, albeit good, is statistically inferior.On the other hand, no statistically significant difference was found when comparing the correlations between Averages 2 and 3; 2 and 4, and 3 and 4.This finding indicates that the correlations of the SDS and Averages 2, 3 and 4, in addition to being superior, are also quite similar; therefore, it is impossible to infer which of these three averages is the best predictor.Nevertheless, these results allow us to conclude that the frequencies 3000 Hz and 4000 Hz are important for speech recognition, since the correlation of Average 1 was lower with the SDS, and statistically different as compared with the other three averages.To corroborate these findings, the MSE (Table 2) was lower for Average 3, followed by Average 2-albeit with quite similar values.This shows that both averages, 2 and 3, have a greater predictive value regarding the SDS, i.e., the more severe the impairment at 3000 Hz and 4000 Hz, the worse the performance in the SDS test.This finding is explained in the literature 7,9 which indicates that 60% of speech intelligibility relies on frequencies above 1000 Hz.

Figure 1 -
Figure 1 -Means for the pure-tone thresholds by frequency of 482 ears

-
p-value for the SDS -p-value for the SRT Average 1-mean of frequencies 500 Hz, 1000 Hz and 2000 Hz.Average 2-mean of frequencies 500 Hz, 1000 Hz, 2000 Hz and 4000 Hz.Average 3-mean of frequencies 500 Hz, 1000 Hz, 2000 Hz and 3000 Hz.Average 4-mean of frequencies 500 Hz, 1000 Hz, 2000 Hz, 3000 Hz and 4000 Hz.Hypothesis testing of the equality of the correlations of Spearman's coefficient of correlation.Level of significance: p<0.05.

Figure 2 -
Figure 2 -Correlation matrix with hypothesis testing p-values comparing the significance of the correlations for the SRT and SDS 15

Table 1 -Descriptive analysis of the pure-tone frequencies, calculated means, mean SRT and SDS in the study sample
* -values in %

Table 2 -Correlations of the individual frequencies 500 Hz, 1000 Hz, 2000 Hz, 3000 Hz and 4000 Hz, Average 1, Average 2, Average 3 and Average 4 with the SRT and the SDS
whether the correlations found in Table2are statistically different, both for the SRT and SDS.In Figure2, a matrix with hypothesis testing p-values is presented.Using that matrix, it is possible +best variable as a predictor of correlation Spearman's coefficient of correlation test Mean Squared Error test to assess