Predicting Hearing Aid Outcomes Using Machine Learning

According to the World Health Organization (WHO) [1], by 2050, 2.5 billion people will suffer from hearing loss, and 700 million of them will need hearing aids (HAs). Currently, 5–6% of the world’s population (430 million people) require treatment for what is known as disabling hearing loss. Disabling hearing loss is defined as a hearing loss of more than 35 dB in the better ear. The impact of hearing loss [2] can be far-reaching, affecting social and emotional well-being, communication, mental health, and working life. In addition to the individual impact, hearing loss can be a burden on third parties such as relatives and caregivers [3]. There is also growing evidence of a link between hearing loss and cognitive function and neurocognitive disorders such as dementia [4]. Hearing rehabilitation [5] with HAs is the only therapeutic alternative for many people [6]. It is therefore essential to ensure its effectiveness [7, 8]. When a person’s hearing is impaired, their first need is to improve understanding in their everyday environment, both in quiet and in noise. The aims of this study were to measure the effectiveness of HAs in improving understanding in quiet and noise and to investigate the factors that significantly influence these results.

A retrospective cohort analysis was conducted using big data analysis to evaluate HA outcome and identify its predictive factors. In particular, this analysis assessed the impact of HA technology (brand and category level), personalized fitting by the hearing care professional (amplification levels by frequency and binaural loudness balancing), and patient follow-up and daily use of HAs on HAs outcome.

HA centers are an integrated network of hearing care professionals specialized in HA fitting. Between 2018 and 2021, 77,661 patients were fitted in HA centers, using common methodologies and measurement tools that allowed the collection of homogeneous audiological data.

The study is based on the analysis of a large population of HA users who were fitted in HA centers between 2018 and 2021, over a 3-year period. Only subjects who completed the HA trial and wore the HA were included. They received routine clinical care based on a single protocol using common measurement tools and methods, and the study was retrospective. Without HAs, the hearing test conditions were audiograms with headphones on the left and right ear, speech perception in quiet (SPIQ) tests in free field (binaural), speech perception in noise (SPIN) tests in free field (binaural). With HAs, the hearing test conditions were audiograms performed in free field (without headphones) binaural, SPIQ tests performed in free field (binaural), SPIN tests performed in free field (binaural). The data have been anonymized so that it is no longer possible to trace the data back to specific hearing care professionals at the user level. Authorization to use the data for retrospective analysis was covered by the General Data Protection Regulation (GDPR) agreement.

The analysis of HA performance is based on the level of intelligibility in quiet and noise, in both cases comparing aided and unaided results. At the initial screening, the audiological assessment should include the measurement of air conduction thresholds (pure-tone average [PTA] averaged over frequencies of 0.5, 1, 2, and 4 kHz) in both ears. The assessments should be made using a standard audiological setup in a sound-treated room. Stimuli should be presented via headphones, with the better ear masked using in-ear earphones.

First, the tonal audiogram was measured for each ear using headphones. The level of detection for the frequency is 125, 250, 500, 750, 1,000, 1,500, 2,000, 3,000, 4,000, 6,000, 8,000 Hz. The degree and type of hearing loss were classified using the four-frequency (500 Hz, 1,000 Hz, 2,000 Hz, and 4,000 Hz) PTA thresholds (PTA4) and the World Health Organization criteria [1].

Only patients with symmetric audiometry were included in the study. For patient selection, the analysis was based on the Bureau International d’AudioPhonologie (BIAP) definition of asymmetry. It should be noted that there are several definitions in audiology circles. Some definitions compare the PTA per ear, while our definition is worded as follows “a person’s hearing is said to be asymmetrical if there is a frequency among those in the PTA (500 Hz, 1,000 Hz, 2,000 Hz, 4,000 Hz) where the difference in detection threshold between the two is greater than or equal to 15 dB.” In the database, based on this definition, 34% of people have asymmetrical hearing loss, which corresponds to the values published in the literature [9]. For the remainder of the study, only people with symmetrical hearing loss are kept in the database to limit the factors that aggravate hearing loss and to maintain a homogeneous population. This is also by design, as some binaural tests may not reflect the true loss in someone with an asymmetrical loss. To obtain the tonal gain of amplification, the free field tonal audiogram was obtained with and without HAs in the same way, using a single loudspeaker in front of the patient.

The SPIQ test is a quiet speech audiometry test that involves two main measurements: the first measurement expressed in dB uses an adaptive procedure to quickly determine the speech recognition threshold (SRT), i.e., the level at which 50% of words are repeated correctly. The second, more conventional measurement is used to find the maximum level of intelligibility. It can only be performed after the SRT has been calculated. This value is expressed as a percentage. The second measurement corresponds to the proportion of words understood, usually out of 10, when 15 dB is added to the SRT. The speech material used is the Lafon dissyllabic lists. This test was carried out in free field conditions, i.e., in bilateral conditions, with and without HAs.

The SPIN test indicates the level at which the patient understands 50% of simple sentences in noise. SPIN is an adaptive free field test using cue sentences in French. It measures the signal/noise ratio (SNR) required to achieve 50% intelligibility. The noise (ICRA type) is fixed at 60 dB. Both signal and noise come from a single loudspeaker in front of the patient, which is the most difficult acoustic condition for HA treatment, as there is no effect of HA directivity and binaural unmasking [10]. Variations in SPIN therefore represent the minimum improvement in noise for the patient but make the test fully reproducible (regardless of the acoustic environment). The result is an SNR value expressed in dB. Full details of the SPIN test used in this study have been published previously [11].

When considering a binaural fitting, the first consideration is likely to be the level of stimulation. The most common fitting goal in this context is “loudness balancing” [12]. Loudness balancing is probably the most plausible goal for optimizing speech understanding in noise. In the desirable case of binaural fusion, there is no isolated left or right loudness, and the fitting goal is rather a centralized perception. Today’s HAs contain multichannel adaptive directional microphones capable of positioning the minimum pick-up in the azimuth of the noise. With these new systems, all localization systems with multiple loudspeakers have become ineffective and unproductive under HA-wearing conditions. This test allows binaural loudness balancing to be carried out by adjusting the gain of the right and left HAs, after testing with a loudspeaker in front of the patient, by generating narrowband noise (for 250, 500, 1,000, 1,500, 2,000, 3,000, 4,000, 6,000, and 8,000 Hz) for a moderate sound (60 dB) and after a loud sound (80 dB) and at least one low sound approximately 10 dB below the aided level of pure-tone audiometry at 2,000 Hz. This test is used to balance the sound between the left and right HAs. It uses the same loudspeaker as previously used for the SPIQ and SPIN tests in free field conditions.

It provides true stereo acoustics at all input levels, optimizing the brain’s unmasking and denoising functions. It is performed with all the signal processing used in HAs. Studies have shown that the slope of the loudness curve just after threshold can be as high as 15/1, whereas the slope of the curve for a normal hearing person is 1/1: the loudness of a hearing-impaired person a few dB above threshold can vary greatly, from soft to very loud.

Fine-tune the balance. In reality, it is unrealistic to expect a precise binaural loudness balance to be set on the day of delivery and to remain constant over time (unless it is renewed). The patient has to adapt to the gain, the central gain has to adapt, and habituation has to take place.

The software allows you to note the positions and make corrections. The test is repeated until balance is achieved. The balance is adjusted for all gains (G 40 dB, G 60 dB, G 80 dB or soft, medium, and loud sounds) at the frequency in question, dropping 1 dB on the localized side. This test is often accurate to the nearest dB.

Obviously, corrections will be much smaller if the initial targeting has been done using a method that takes into account the patient’s level of discomfort. This procedure is often more productive in terms of improving understanding in noise than activating the HA’s anti-noise systems. This is done at the initial HA fitting and at follow-up appointments.

To ensure a better homogeneity of the etiology of the hearing loss, the selected patients were therefore focused on presbycusis etiology. Therefore, only patients over 50 years of age with at least one pure-tone audiogram with and without HAs were included in the study. The age range was 50 to 90 years. A total of 60,835 and 77,661 subjects (Table 1) were included in the analysis of speech in quiet and speech in noise data with and without HAs.

All patients have a fitted audiogram and a fitted SPIQ or SPIN test. For each patient in the database, we kept the last fitted audiogram, the last fitted SPIN test, and the last fitted SPIQ test performed. We excluded from the study all patients who had no value for any of the fitted tests (audiogram, SPIN, SPIQ) and those who had no unfitted audiogram. The study will also look at other values: age of the patients at the time of the first fitting, gender, HA technology (brand or manufacturer, level of technology), device setting; adherence to HAs: time spent wearing the device during the first month; history (Boolean values): noise exposure, family, diabetes, ontological, presence of tinnitus (Boolean value), and a scale to determine the intensity and level of discomfort of tinnitus, and hearing care professional fitting accuracy (amplification and stereo balance).

Big data analysis of HA outcome was made in two steps:

• 1

  Identification of predictive factors made by eXtreme Gradient Boosting (XGBoost) [13] machine learning method. This decision tree-based algorithm utilizes the boosting technique for training and can deal with missing values. A traditional 80% training and 20% testing framework was implemented, along with grid search using 5-fold cross-validation. The best set of hyperparameters resulted in a 3-dB 95th percentile of absolute error for SPIQ test and 1 dB for SPIN test.

• 2

  Evaluation of isolated impact of the different variables through SHapley Additive exPlanations (SHAP) Value analysis [14] and interpretation of model predictions using the SHAP values framework [15]. Derived from game theory, SHAP values provide the individual impact of each feature on predictions, independent of the other inputs. For each sample, the cumulative sum of these impact values represents the difference between the actual prediction (f(x)) and the average prediction of the model (E(f(x))). By averaging the SHAP absolute values across the entire dataset, we obtained the mean impact of each feature, enabling their ranking and deriving their importance.

We applied the SHAP method to explain the best-performing predictive model. Feature ranking was obtained by computing SHAP values. The features were ordered by the mean absolute value of the SHAP values for each feature.

The graphic representation gives a color depending on the intensity of the hearing loss. Blue represents light to medium losses, and red represents medium to severe losses according to the color indicator.

There are three types of missing data.

• -

  Missing completely at random when the events leading to the disappearance of a particular data item are independent of both observable variables and unobservable parameters of interest, and yet occur randomly.

• -

  Randomly missing data when the missing observations are not random but can be fully explained by variables where complete information exists.

• -

  Predictably missing by omission (PMO), also known as nonresponse or non-ignorable data, is data that are neither randomly missing data nor PMO. In this sense, the value of the missing variable is linked to the reason why it is missing.

Only the first type of missing data does not bias the analysis. In reality, and in this database, most of the missing data in the database are PMOs. For example, above an average loss of 70 dB, hearing care specialists no longer give either the SPIQ test or the SPIN test, as the loudspeaker capacities are too low in view of the patient’s loss. To avoid bias as far as possible, we chose to keep the population as large as possible: deleting the population missing a test could have introduced bias into the study. As soon as a patient had completed at least one of the tests, he or she was included in the general statistics study. The advantage of XGBoost is that there is no need to replace missing data with more or less arbitrary data. It will consider it as data and deduce its value. To measure the impact of antecedents on hearing, we systematically compare patients who responded to the questionnaires only. The same applies to the tests, as this enables us to measure the impact and compare two “comparable” populations. In the database we used, 80% of patients had a non-instrumented audiogram, which was the only exclusion criterion in our study. After exclusion, gender, age, the level of technology, and provider were present for 100% of patients. Finally, all patients who had no value for the fitted tests (audiogram, SPIN, SPIQ) and those who had no unfitted audiogram were included in the study.

The patient population is equally divided between men and women. The average age and hearing loss are 73 years and 50 dB HL, respectively. According to the Bureau International d’AudioPhonologie (BIAP) classification, the distribution of the population is as follows: mild (0–39.9 dB) 18%; moderate level 1 (40–55 dB) 45%; moderate level 2 (56–70 dB) 28%; severe and profound (71 dB and more) 9%. After fitting, the average PTA is 29 dB, giving a gain of 21 dB.

First, for speech in quiet shown in Figure 1 on the right, the average unaided SRT of the cohort is 51 dB SPL. With HAs, the average SRT is then 40 dB. The gain is 11 dB. A total of 75% of the population achieved an SRT of less than 46 dB after fitting, which corresponds to the level of a whispered voice. Overall, 99% of patients regain an SRT below 60 dB, which is the level of normal speech.

Second, for speech in noise shown in Figure 1 on the left, the average unaided SNR for all patients is 5.9 dB SNR. After fitting, the average SNR is 2.8 dB SNR, resulting in an average gain of 3.1 dB. Hence, 75% of the population gained at least 2 dB SNR, a significant improvement.

To understand the role of gain in the speech in quiet and speech in noise results, the SHAP values shown in Table 2 lead to the same main contributor for both tests with the gain values around 2,000 Hz at an intensity level in the range of 15–25 dB that is decisive for the results. The second contributor is slightly different with the gain values around 1,000 Hz for speech in quiet, whereas in noise it is the gain values around 4,000 Hz that maximize the result.

In order to classify the level of technology shown in Figure 2, the data are classified according to the 4 levels of HA technology. Level 2 corresponds to basic technology with a minimum level of HA performance, with no restrictions on the output level of the HA (level 2 in the international classification). Higher technology levels are divided into 3 technology levels from 3 to 5.

The higher the technology level, the greater the improvement in understanding both in quiet and in noise. This is particularly true for SPIN, where the improvement in HA results applies to all levels of HA and all degrees of hearing loss.

For SPIQ, technology level has less of an effect for mild to moderate 1 hearing losses than for moderate 2 to severe hearing losses. This is particularly significant when comparing level 5 technologies with other levels.

The importance of binaural loudness balancing is particularly evident for the SPIN [16] with range 5 at the SHAP value. The effect of binaural loudness balancing is clearly shown in Figure 3.

The population was divided into two parts using a Boolean indicator: yes or no. The yes condition means that the patient performed loudness balancing tests with their HAs during the fitting appointments. The “no condition” means that the patient did not perform a loudness balance test. Binaural loudness balancing always leads to a better HA result in both quiet and noise, and the effect is particularly important for hearing losses above 50 dB.

The final gain on the speech in quiet and speech in noise tests in Figure 4 shows the effect of adherence to therapy on the improvement in intelligibility of the patients both in quiet and in noise. The results are presented by level of hearing loss (unaided): from 40 to –49; 50 to 59; 60 and above. A significant improvement in results is seen when the HA is worn for more than 9 h a day. For understanding in noise, the effect of daily use is particularly important for patients with hearing losses above 60 dB.

There is no comparable analysis of HA outcomes based on such a large cohort of 77,661 patients in the literature [17‒19]. This study shows that the use of big data [20] for retrospective analysis is an important method to evaluate HA outcomes.

HAs provide significant results for the vast majority of patients, in terms of understanding both in quiet and in noise. The average level of improvement per patient is also very significant (11 dB for SRT; 3.1 dB for SNR). Although the average results are very significant, there is still a significant variation. This suggests that there are different types of audiological profiles for which results and predictive factors may vary. To take the analysis a step further, we would need to predefine clusters of patients based on audiological profiles (supervised approach) and perform the analysis for each profile separately.

Technology [21] affects patient outcomes in both quiet and noisy environments. The higher the level of technology, the greater the impact. Unlike a previous publication based on patient questionnaires [22], here the impact is shown in terms of improved understanding in quiet and especially in noise. This effect is more important for more severe hearing losses. Another study [23] suggests that patients appreciate the higher level of technology in the complex auditory situations of everyday life.

There are several publications [24] on the effects of binaural loudness compensation on speech understanding [25], especially in noise. This study shows that this effect is significant, particularly for understanding in noise, even when only patients with symmetric hearing loss are included. The positive effect of binaural loudness balancing on the outcome of HAs is true for all patients, especially for those with a hearing loss of more than 60 dB. It would be particularly interesting to perform the same analysis for patients with asymmetrical hearing loss.

As has been extensively reported in the literature [26, 27], there is an impact on patient adherence to HA use and HA outcomes. This study shows that there are clearly 2 groups of patients, with a segmentation point at 9 h per day of HA use. Patients who use their HA for at least 9 h per day have a better outcome in quiet and especially in noise. In noise, this is particularly true when the hearing loss is greater than 60 dB.

The effectiveness of HA fitting in improving speech perception in both quiet and noise was measured and proven by the study. HA outcomes are significant in terms of both the level of improvement per patient and the percentage of patients improved.

The study also provides a comprehensive understanding of the factors that determine the outcome of HA fitting. In particular, the analysis shows that the hearing care professional has an impact on the outcome through the choice of the most appropriate technology, the use of personalized fittings, and the subsequent adherence to therapy. In particular, binaural loudness balancing was shown to be systematically beneficial.

However, global results must be interpreted with caution in such a heterogeneous population. The results of the study would need to be validated and refined by an approach using clusters of patients with similar audiological profiles. This approach would allow the development of an understanding of the determinants of HA outcome relevant to a specific patient profile and therefore applicable in practice by the hearing care professional to optimize HA outcome.

T.L. thanks Jean-Marc Alliot, Institut de Recherche en Informatique de Toulouse (IRIT). All authors are grateful to the audiologists who collect all the data on a daily basis in the clinical practice with the patients.

All procedures in this study were in accordance with the ethical standards of the Institutional Research Committee and with the Helsinki Declaration. The study adhered to ethical guidelines, and written informed consent was obtained from all participants involved in the study. This study protocol was reviewed and approved by the Institutional Review Board CNIL, Approval No. 22144039 v 0, dated June 19, 2019. Patient consent is obtained by informing the patient in accordance with the MR004 reference Methodology. Written informed consent was obtained from the patient for participation to this study.

P.R., C.B., E.B-.M., and F.A. are employees of Amplifon Company. T.L. had a fixed-term contract with Amplifon during the analyzed phase of the study. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

This study was funded by Amplifon France.

E.B.-.M. and F.A. designed the study and collected the data. T.L. performed the machine learning and statistical analysis. S.M. supervised the machine learning and statistical analysis. All authors reviewed the data from all sites and performed the interpretative analysis. P.R. and C.B. drafted the manuscript. B.F. is responsible for overall content as guarantor. All authors discussed the results and implications, commented on the manuscript at all stages, contributed to the writing of the article, and approved the version submitted.

The data that support the findings of this study are not publicly available due to their containing information that could compromise the privacy of research participants. According to the authors’ data use agreement, the script used in this study may only be shared on the basis of a reasoned request. Requests for access to these datasets should be addressed to C.B. at [email protected].