=Paper=
{{Paper
|id=Vol-2439/4-paginated
|storemode=property
|title=Rethinking hearing aids as recommender systems
|pdfUrl=https://ceur-ws.org/Vol-2439/4-paginated.pdf
|volume=Vol-2439
|authors=Alessandro Pasta,Michael Kai Petersen,Kasper Juul Jensen,Jakob Eg Larsen
|dblpUrl=https://dblp.org/rec/conf/recsys/PastaPJL19
}}
==Rethinking hearing aids as recommender systems==
https://ceur-ws.org/Vol-2439/4-paginated.pdf
Rethinking Hearing Aids as Recommender Systems
Alessandro Pasta Michael Kai Petersen
Technical University of Denmark Eriksholm Research Centre
Kongens Lyngby, Denmark Snekkersten, Denmark
alpas@dtu.dk mkpe@eriksholm.com
Kasper Juul Jensen Jakob Eg Larsen
Oticon A/S Technical University of Denmark
Smørum, Denmark Kongens Lyngby, Denmark
kjen@oticon.com jaeg@dtu.dk
ABSTRACT hearing aids is rarely utilized as devices are frequently dispensed
The introduction of internet-connected hearing aids constitutes a with a “one size fits all” medium setting, which does not reflect
paradigm shift in hearing healthcare, as the device can now po- the varying needs of users in real-world listening scenarios. The
tentially be complemented with smartphone apps that model the recent introduction of internet-connected hearing aids represents
surrounding environment in order to recommend the optimal set- a paradigm shift in hearing healthcare, as the device might now be
tings in a given context and situation. However, rethinking hearing complemented with smartphone apps that model the surrounding
aids as context-aware recommender systems poses some challenges. environment in order to recommend the optimal settings in a given
In this paper, we address them by gathering the preferences of seven context.
participants in real-world listening environments. Exploring an au- Whereas a traditional recommender system is built based on data
diological design space, the participants sequentially optimize three records of the form < user,item,rating > and may apply collaborative
audiological parameters which are subsequently combined into a filtering to suggest, for instance, new items based on items pre-
personalized device configuration. We blindly compare this configu- viously purchased and their features, recommending the optimal
ration against settings personalized in a standard clinical workflow hearing aid settings in a given context remains highly complex.
based on questions and pre-recorded sound samples, and we find Rethinking hearing aids as recommender systems, different device
that six out of seven participants prefer the device settings learned configurations could be interpreted as items to be recommended
in real-world listening environments. to the user based on previously expressed preferences as well as
preferences expressed by similar users in similar contexts. In this
CCS CONCEPTS framework, information about the sound environment and user
intents in different soundscapes could be treated as contextual in-
• Information systems → Personalization; Recommender sys-
formation to be incorporated in the recommendation, building a
tems; • Human-centered computing → Ambient intelligence;
context-aware recommender system based on data records of the
User centered design.
form < user,item,context,rating > [1]. However, addressing some
KEYWORDS challenges related to the four aforementioned data types is essen-
tial to make it possible to build an effective context-aware recom-
Personalization, recommender systems, hearing healthcare, hearing mender system in the near future. In this paper, we discuss the main
aids challenges posed when rethinking hearing aids as recommender
ACM Reference Format: systems and we address them in an experiment conducted with
Alessandro Pasta, Michael Kai Petersen, Kasper Juul Jensen, and Jakob Eg seven hearing aid users.
Larsen. 2019. Rethinking Hearing Aids as Recommender Systems. In Pro-
ceedings of the 4th International Workshop on Health Recommender Systems
co-located with 13th ACM Conference on Recommender Systems (HealthRec-
Sys’19), Copenhagen, Denmark, September 20, 2019, 7 pages. 1.1 Rating
In order to be able to precisely and accurately recommend optimal
1 INTRODUCTION
device settings in every situation, gathering relevant user prefer-
Despite decades of research and development, hearing aids still fail ences (expressed as ratings) is essential. However, learning user
to restore normal auditory perception as they mainly address the preferences poses some challenges. Firstly, the device settings re-
lack of amplification due to loss of hair cells in the cochlea [16], flect a highly complex audiological design space involving multiple
rather than compensating for the resulting distortion of neural interacting parameters, such as beamforming, noise reduction, com-
activity patterns in the brain [22]. However, the full potential of pression and frequency shaping of gain. It is important to explore
the different parameters, in order not to disregard some parameters
HealthRecSys’19, September 20, 2019, Copenhagen, Denmark that might have relevant implications for the user listening experi-
© 2019 Copyright for the individual papers remains with the authors. Use permitted ence, and to identify which parameters in an audiological design
under Creative Commons License Attribution 4.0 International (CC BY 4.0). This
volume is published and copyrighted by its editors. space [10] define user preferences in a given context. Secondly, the
preferred device settings depend on the human perception of the
11
HealthRecSys’19, September 20, 2019, Copenhagen, Denmark Alessandro Pasta, Michael Kai Petersen, Kasper Juul Jensen, and Jakob Eg Larsen
listening experience and it is therefore difficult to represent the per- have similar preferences due to individual differences in their sen-
ceptual objective using an equation solely calculated by computers sorineural processing [16, 22]. Therefore, at least in the first phase,
[21]. Having to rely on user feedback, it is important to limit the we need to ask the same user many times about her preferences,
complexity of the interface, to make the interaction as effective as until her optimal configuration is found. Furthermore, in order to
possible. Thirdly, capturing user preferences in multiple real-world optimize the device in different listening scenarios, we need to ask
situations not only guarantees that the situations are relevant and the same user to move in the same design space multiple times.
representative of what the user will experience in the future, but Altering the one-dimensional slider at every step of the evaluation
it also allows the user to test the settings with a precise and real procedure might make the task difficult, since the user would not
intent in mind. However, this increases the complexity of the task, know the trajectory defined by the new slider. We believe that
since the real-world environment is constantly changing and a user decoupling the parameters and allowing users to manipulate one
might explore the design space while performing other actions (e.g. parameter at a time, moving in a one-dimensional space that is
conversing). clearly understood, would allow them to better predict the effects
A traditional approach to find the best parameter combination of their actions and hence more effectively assess their preferences.
(i.e. the best device configuration) is parameter tweaking, which
consists in acting on a set of (either continuous or discrete) param- 1.2 Item
eters to optimize them. Similarly to enhancing a photograph by In order to enhance the hearing aid user experience, it is important
manipulating sliders defining brightness, saturation and contrast to appropriately select the parameters that define the hearing aid
[21], the hearing aid user could control her listening experience configurations evaluated by users. Indeed, not only should the
by tweaking the parameters that define the design space and find parameters have a relevant impact on the user listening experience,
the optimal settings in different listening scenarios. However, this but the different levels of the parameters should also be discernible
method can be tedious when the user is moving in a complex de- by untrained users. Three parameters have been demonstrated to
sign space defined by parameters that interact among each other be particularly important for the experience of hearing impaired
[13]. One frequently used method to simplify the task of gather- users:
ing preferences is pairwise comparison, which consists in making
users select between two contrasting examples. A limitation of (1) Noise reduction and directionality. Noise reduction reduces
this approach is efficiency, given that a single choice between two the effort associated with speech recognition, as indicated by
examples provides limited information and many iterations are pupil dilation measurements, an index of processing effort
required to obtain the preferred configuration. Based on pairwise [23]. By allowing speedier word identification, noise reduc-
comparisons, an active learning algorithm may apply Bayesian tion also facilitates cognitive processing and thereby frees
optimization [2] to automatically reduce the number of examples up working memory capacity in the brain [18]. Moreover,
needed to capture the preferences [3], assuming that the samples fast-acting noise reduction proved to increase recognition
selected for comparison capture all parameters across the domain. performances and reduce peak pupil dilation compared to
Alternatively, one might decompose the entire problem into a se- slow-acting noise reduction [23]. Given that the ability of
quence of unique one-dimensional slider manipulation tasks. As users to understand speech in noisy environments may vary
exemplified by Koyama et al. [13], the color of photographs can be by up to 15 dB [4], it is essential to be able to individualize
enhanced by proposing users a sequence of tasks. At every step, the the threshold levels for the activation of noise reduction.
method determines the one-dimensional slider that can most effi- (2) Brightness. While a lot of research has been focused on adapt-
ciently lead to the best parameter set in a multi-dimensional design ing the frequency-specific amplification which compensates
space defined by brightness, contrast and saturation. Compared to for a hearing loss based on optimized rationales like VAC+
pairwise comparison tasks, the single-slider method makes it possi- [5], rationales still reflect average preferences across a popu-
ble to obtain richer information at every iteration and accelerates lation rather than individual ones. Several studies indicate
the convergence of the optimization. that some users may benefit from increasing high-frequency
Inspired by the latter approach we likewise formulate the learn- gain in order to enhance speech intelligibility [11, 12].
ing of audiological preferences in a given listening scenario as an (3) Soft gain. The perception of soft sounds varies largely among
optimization problem: individuals. Hearing aid users with similar hearing losses
can perceive sounds close to the hearing threshold as being
z = arg max f (x) soft or relatively loud. Thus, proposing a medium setting for
x ∈X amplification of soft sounds may seem right when averag-
where x defines parameters related to beamforming, attenuation, ing across a population, but would not be representative of
noise reduction, compression, and frequency shaping of gain in an the large differences in loudness perception found among
audiological design space X [10] and the global optimum of the individual users [17]. For this reason, modern hearing aids
function f : X → ℜ returns values defining the preferred hearing provide the opportunity to fine-tune the soft gain by acting
aid settings in a given listening scenario. on a compression threshold trimmer [14].
However, while it remains sensible to assume that individual ad- Taking a naive approach, treating each parameter independently,
justments would converge when crowdsourcing (i.e. asking crowd the preferences could subsequently be summed up in a general hear-
workers to complete the tasks independently) the task of enhancing ing aid setting, by simply applying the most frequently preferred
an image [13], it is less likely that hearing impaired users would values along each audiological parameter.
12
Rethinking Hearing Aids as Recommender Systems HealthRecSys’19, September 20, 2019, Copenhagen, Denmark
1.3 User 0
Hearing aids are often fitted based on a pure tone audiometry, a
10 Left
test used to identify the hearing threshold of users. However, as
mentioned above, users perceive the sounds differently and might Right
20
Hearing Threshold (dB HL)
benefit from a fully personalized hearing aid configuration. For this
reason, it is essential to fully understand what drives user prefer- 30
ences and which is the relative importance of users’ characteristics 40
and context. It is interesting to analyse whether users exhibit similar
preferences when optimizing the hearing aids in several real-world 50
environments and whether they result into similar configurations. 60
70
1.4 Context 80
Users often prefer to switch between highly contrasting settings de- 90
pending on the context [11]. It has been shown that a context-aware
hearing aid needs to combine different contextual parameters, such 100
100
125
250
500
750
1000
1500
2000
3000
4000
6000
8000
9000
as location, motion, and soundscape information inferred by audi-
tory measures (e.g. sound pressure level, noise floor, modulation
envelope, modulation index, signal-to-noise ratio) [12]. However, Frequency (Hz)
these contextual parameters might fail to capture the audiological
intent of the user, which depends not only on the characteristics of Figure 1: Average hearing threshold (i.e. the sound level be-
the sound environment but also on the situation the user is in. For low which a person’s ear is unable to detect any sound [7])
this reason, in addition to retrieving the characteristics of the sound levels for the 7 participants. The participants had a hearing
environment and the preferred device settings, it is also important loss ranging from mild to moderately severe. Error bars in-
to capture the contextual intents of users in the varying listening dicate ±1 standard deviation of the hearing thresholds.
scenarios. Contextual information, in this exploratory phase, can be
explicitly obtained by directly asking the user to define the situation
she is in. However, in the future, to enable an automatic adaptation
to the needs of users in real-world environments, relevant contex-
tual information will need to be inferred using a predictive model
2.3 Procedure
that classifies the surrounding environment. The experiment was divided into four weeks. As shown in Table 1,
the first three weeks were devoted to optimizing the three audio-
logical parameters, one at a time. Each of the first three weeks, the
participants were fitted with four levels of the respective parameter,
2 METHOD
while the other two parameters were kept neutral at a default level.
2.1 Participants For instance, in week 1, each participant could select between four
Seven participants (6 men and 1 woman), from a screened popula- levels of noise reduction and directionality. The participants were
tion provided by Eriksholm Research Centre, participated in the instructed to compare, using a smartphone app, the four levels of
study. Their average age was 58.3 years (std. 12 years). Five of them the parameter in different situations during their daily life and to
were working, while two were retired. They were suffering from a report their preference. To ensure that the participants would eval-
binaural hearing loss ranging from mild to moderately severe, as uate the different levels in relevant listening situations and when
classified by the American Speech-Language-Hearing Association motivated to optimize their device, they were instructed to perform
[6]. The average hearing threshold levels are shown in Figure 1. the task on a voluntary basis. Moreover, every time they reported
They were all experienced hearing aid users, ranging from 5 to 20 their preference, the participants were asked to specify:
years of experience with hearing aids. All test subjects received
• The environment they were in (e.g. office, restaurant, public
information about the study and signed an informed consent before
space outdoor). Different environments are characterised
the beginning of the experiment.
by different soundscapes and pose disparate challenges for
hearing aid users.
• Their motion state (e.g. stationary, walking, driving). Mo-
2.2 Apparatus tion tells more about the activity conducted by the person,
The participants were fitted according to their individual hearing but may also mark the transition to a different activity or
loss with a pair of Oticon Opn S 1 miniRITE [8]. All had iPhones environment [9].
with iOS 12 installed and additionally downloaded a custom smart- • Their audiological intent (e.g. conversation, work meeting,
phone app connected to the hearing aids via Bluetooth. The app watching TV, listening to music, ignoring speech). Comple-
enabled collecting data about the audiological preferences and the menting the contextual information by gathering the intent
corresponding context. of the participants in the specific situation might provide a
13
HealthRecSys’19, September 20, 2019, Copenhagen, Denmark Alessandro Pasta, Michael Kai Petersen, Kasper Juul Jensen, and Jakob Eg Larsen
Table 1: Study timeline
Noise Reduction and Directionality (n=94)
Brightness (n=98)
Week Activity Soft Gain (n=103)
W. 1 Optimization of noise reduction and directionality
5
W. 2 Optimization of brightness (amplification of high-frequency
4
Usefulness
sounds)
W. 3 Optimization of soft gain (amplification of soft sounds) 3
W. 4 Final test of preference
2
deeper insight into how the different audiological parameters 1
A B C D E F G
help them in coping with different sounds.
Participant
• The usefulness of the parameter in the specific situation (on a
scale ranging from 1 to 5). This evaluation is important not
only to understand the relative importance of each prefer- Figure 2: Average perceived usefulness of three parameters
ence, but also to assess the perceived benefit of the parameter (noise reduction and directionality, brightness, soft sounds).
in diverse situations. Brightness is perceived to be the most useful parameter.
Noise reduction and directionality tends to be perceived as
The fourth week each participant compared two different device
the least useful parameter.
configurations in a blind test:
• An individually personalized configuration combining the
most frequently selected preferences of the three audiologi- recorded in situations where the usefulness of the parameter is
cal parameters gathered in real-world listening environments rated higher than two out of five are considered.
during the previous three weeks. Firstly, the results indicate that the participants have widely dif-
• A configuration personalized in a standard clinical work- ferent audiological preferences, rather than converging towards
flow based on questions and on pairwise comparisons of a shared optimal value. As the participants are ordered by age (A
pre-recorded sound samples capturing different listening being the youngest), there seem, nevertheless, to be some com-
scenarios including, for instance, speech with varying levels mon tendencies among younger or older participants across all
of background noise. parameters.
Secondly, most participants are not searching for a single op-
The participants were instructed to compare the two personalized
timum but select different values within each parameter. When
configurations in different listening situations throughout the day
adjusting the perceived brightness (Figure 4), six participants out
and report their preference, while also labeling the context. At the
of seven prefer, most of the time, the two highest levels along this
end of the week, the participants were asked to select the configu-
parameter. Thirdly, the participants frequently prefer highly con-
ration they preferred.
trasting values within each parameter, depending on the context.
3 RESULTS
During the four weeks of test, the participants actively interacted Noise Reduction and Directionality
with their devices, changing the hearing aid settings, overall, 4328 100%
times (i.e. the level of the parameter during the first three weeks
80% Level 4
or the final configuration during the last week) and submitting 406
Preference
60% Level 3
preferences. On average, the participants tried the different hearing
aid settings 11 times before submitting a preference. Although one Level 2
40%
parameter affects the perception of the others, isolating them al- Level 1
20%
lows to analyse their perceived impact on the listening experience.
As illustrated in Figure 2, the brightness parameter was on aver- 0%
A B C D E F G
age rated higher in perceived usefulness. This result is consistent (n=5) (n=9) (n=0) (n=4) (n=13) (n=1) (n=20)
among the seven participants. Conversely, the noise reduction and
Participant
directionality parameter resulted to have the lowest perceived use-
fulness for five participants out of seven. The soft gain parameter
resulted to have an average perceived usefulness between those of Figure 3: Preferences for the 4 levels of noise reduction and
the other two parameters. directionality, which correspond (from level 1 to level 4) to
Recording, together with each preference, the perceived use- increasing directionality settings, increasing levels of noise
fulness of the parameter in the specific situation also allows to reduction in simple and complex environments and earlier
understand how much each parameter contributes to the overall activation of noise reduction [15]. The participants exhib-
setting of the hearing aid. Figures 3, 4, 5 display the preferences of ited different noise reduction and directionality preferences
test participants for different levels of noise reduction and direction- and five of them preferred more than one level in different
ality, brightness, and soft gain, respectively. Only the preferences situations.
14
Rethinking Hearing Aids as Recommender Systems HealthRecSys’19, September 20, 2019, Copenhagen, Denmark
Brightness 4 DISCUSSION
100%
Due to the aging population, the number of people affected by hear-
ing loss will double by 2050 [20] and this will have large implications
80% Level 4 for hearing healthcare. Rethinking hearing aids as recommender
Preference
60% Level 3 systems might enable the implementation of devices that automat-
40%
Level 2 ically learn the preferred settings by actively involving hearing
Level 1 impaired users in the loop. Not only would this enhance the expe-
20% rience of current hearing aid users, but it could also help overcome
0% the growing lack of clinical resources. Personalizing hearing aids by
A B C D E F G integrating audiological domain-specific recommendations might
(n=5) (n=10) (n=7) (n=8) (n=7) (n=12) (n=42)
even make it feasible to provide scalable solutions for the 80% of
Participant hearing impaired users who currently have no access to hearing
healthcare worldwide [19]. The accuracy of the recommendation
Figure 4: Preferences for the 4 levels of brightness, which primarily depends on the ability of the system to gather user pref-
correspond (from level 1 to level 4) to increasing amplifica- erences, while the user explores a highly complex design space. In
tion of high-frequency sounds. The participants exhibited this study, we proposed an approach to effectively optimize the
different brightness preferences and six of them preferred device settings by decoupling three audiological parameters and
more than one level in different situations. allowing the participants to manipulate one parameter at a time,
comparing four discrete levels. The fact that the participants pre-
ferred the hearing aid configuration personalized in real-world
environments suggests that the proposed optimization approach
manages to capture the main individual parameter preferences.
Looking into the individual preferences learned when sequen-
Soft Gain
tially adjusting the three parameters, several aspects stand out. The
100% results suggest that the brightness parameter has the highest per-
80% Level 4 ceived usefulness. This could be due to the fact that enhancing the
Preference
Level 3 gain of high frequencies may increase the contrasts between conso-
60%
nants and as a result improve speech intelligibility. Likewise, it may
Level 2
40% amplify spatial cues reflected from the walls and ceiling, improving
Level 1
20% the localization of sounds and thereby facilitating the separation of
voices. The participants seemed to appreciate a brighter sound when
0%
A B C D E F G
listening to speech or when paying attention to specific sources
(n=6) (n=16) (n=4) (n=0) (n=11) (n=6) (n=37) in a quiet environment. Despite the advances in technology that
Participant reduce the risk of audio feedback and allow the new instruments
to be fitted to target and deliver the optimal gain [8], in some situa-
Figure 5: Preferences for the 4 levels of soft gain, which cor- tions most of the participants seemed to benefit from even more
respond (from level 1 to level 4) to increasing amplification brightness. Conversely, users might prefer a more round sound in
of soft sounds, thus increasing dynamic range compression noisy situations or when they want to detach themselves.
[14]. The participants exhibited different soft gain prefer- Adjusting the noise reduction and directionality parameter is per-
ences and five of them preferred more than one level in dif- ceived as having the lowest usefulness. Essentially, this parameter
ferent situations. defines how ambient sounds coming from the sides and from behind
are attenuated, while still amplifying signals with speech character-
istics. Although the benefits of directionality and noise reduction
are proven, our results indicate that users find it more difficult to
In order to combine the sequentially learned preferences, we differentiate the levels of this parameter if the ambient noise level is
summed up the most frequently chosen values along each param- not sufficiently challenging. The four levels of the parameter mainly
eter into a single hearing aid configuration. For each participant, affect the threshold for when the device should begin to attenuate
we subsequently compared it against individually personalized set- ambient sounds. However, these elements of signal processing are
tings configured in a standard clinical workflow based on questions partly triggered automatically based on how noisy the environment
and pre-recorded sound samples. After the fourth week, six out of is. Therefore, in some situations, changing the attenuation thresh-
seven participants responded they appreciated having more than olds (i.e. the parameter levels) might not make a difference. Thus,
one general hearing aid setting, as they used both configurations in users may feel less empowered to adjust this parameter. On the
different situations. They also wished to keep both personalized con- other hand, the data also shows that participants actively select the
figurations after the end of the test. However, in a blind comparison lowest level of the parameter (level 1), which provides an immersive
of the two configurations, six out of seven participants preferred omnidirectional experience without attenuation of ambient sounds
the hearing aid settings personalized by sequentially optimizing in simple listening scenarios. This suggests that, in some contexts,
parameters in real-world listening scenarios.
15
HealthRecSys’19, September 20, 2019, Copenhagen, Denmark Alessandro Pasta, Michael Kai Petersen, Kasper Juul Jensen, and Jakob Eg Larsen
users express a need for personalizing the directionality settings to have a different perceived usefulness, differently contributing to
and the activation thresholds of noise reduction. Furthermore, pre- the listening experience of hearing aid users. The seven participants
vious studies have shown that the perception of soft sounds varies exhibited widely different audiological preferences. Furthermore,
largely among individuals. Our results not only confirm that users our results indicate that hearing aid users do not simply explore the
have widely different audiological preferences, but also suggest audiological design space in search of a global optimum. Instead,
they would benefit from a personalized dynamic adaptation of soft most of them select multiple highly contrasting values along each
gain dependent on the context. parameter, depending on the context.
Focusing on the optimization problem in the audiological de-
sign space, some indications can be inferred. The large differences ACKNOWLEDGMENTS
among the participants suggest that, in a first phase, users’ inter- We would like to thank Oticon A/S, Eriksholm Research Centre, and
action is essential to gather individual preferences and thereby Research Clinician Rikke Rossing for providing hardware, access
reach the optimum configuration for each single user. Simplify- to test subjects, clinical approval and clinical resources.
ing the optimization task and offering a clear explanation of the
one-dimensional slider made the process more transparent and
REFERENCES
increased users’ empowerment. Once a recommender system is in
[1] Gediminas Adomavicius and Alexander Tuzhilin. 2011. Context-Aware Rec-
place, this component might also prove useful in enhancing users’ ommender Systems. In Recommender Systems Handbook, F. Ricci, L. Rokach,
trust in the recommendations provided. Moreover, performing the B. Shapira, and P. Kantor (Eds.). Springer, Boston, MA, USA, 217–253. https:
//link.springer.com/chapter/10.1007/978-0-387-85820-3_7
optimization task in real-world environments ensured an accurate [2] Eric Brochu, Vlad M. Cora, and Nando de Freitas. 2010. A Tutorial on Bayesian
assessment and communication of users’ preferences. In the short Optimization of Expensive Cost Functions, with Application to Active User
term, user preferences collected with this approach could flow into Modeling and Hierarchical Reinforcement Learning. CoRR abs/1012.2599 (2010).
[3] Eric Brochu, Nando de Freitas, and Abhijeet Ghosh. 2007. Active Preference
the standard clinical workflow and help hearing care professionals Learning with Discrete Choice Data. In Proceedings of the 20th International
to fine-tune the hearing aids. However, a single static configura- Conference on Neural Information Processing Systems (NIPS ’07). 409–416.
tion, although personalized, might not fully satisfy the user. Our [4] Mead C. Killion. 2002. New Thinking on Hearing in Noise: A Generalized
Articulation Index. Seminars in Hearing 23 (January 2002), 057–076. https:
results indicate that such recommender systems should not simply //doi.org/10.1055/s-2002-24976
model users as a sole set of optimized audiological parameters, be- [5] Susanna L. Callaway and Andreea Micula. 2016. Client Target and Real-ear
Measurements. Technical Report. Oticon A/S, Smørum, Denmark.
cause the preferred configuration varies depending on the context. [6] John Clark. 1981. Uses and Abuses of Hearing Loss Classification. ASHA: a
It is therefore essential for these models to likewise classify the journal of the American Speech-Language-Hearing Association 23 (August 1981),
sound environment and motion state in order to infer the intents 493–500.
[7] European Commission. 2019. Glossary: Hearing threshold. Retrieved August
of the user. Being fully aware of the intent, by automatically la- 16, 2019 from https://ec.europa.eu/health/scientific_committees/opinions_
beling it, would add further value to the collected preferences and layman/en/hearing-loss-personal-music-player-mp3/glossary/ghi/hearing-
would allow to ask for user feedback in specific situations. That threshold.htm
[8] Josefine J. Jensen. 2019. Oticon Opn S Clinical Evidence. Technical Report. Oticon
would make it feasible to verify hypotheses based on previous data, A/S, Smørum, Denmark.
and progressively optimize several device configurations for differ- [9] Benjamin Johansen, Maciej J. Korzepa, Michael K. Petersen, Niels H. Pontoppi-
dan, and Jakob E. Larsen. 2018. Inferring User Intents from Motion in Hearing
ent real-world listening scenarios. Once some configurations are Healthcare. In Proceedings of the 2018 ACM International Joint Conference and
learned, the hearing aids could automatically recommend them in 2018 International Symposium. https://doi.org/10.1145/3267305.3267683
specific situations and, by monitoring users’ behavior, continuously [10] Benjamin Johansen, Michael K. Petersen, Niels H. Pontoppidan, and Jakob E.
Larsen. 2019. Modelling User Utterances as Intents in an Audiological Design
calibrate to the preference of the user. Space. In Workshop on Computational Modeling in Human-Computer Interaction
(CHI 2019).
[11] Benjamin Johansen, Michael K. Petersen, Maciej J. Korzepa, Jan Larsen, Niels H.
5 CONCLUSION Pontoppidan, and Jakob E. Larsen. 2018. Personalizing the Fitting of Hearing Aids
by Learning Contextual Preferences From Internet of Things Data. Computers 7,
Internet-connected hearing aids open the opportunity for truly 1 (2018). https://doi.org/10.3390/computers7010001
[12] Maciej J. Korzepa, Benjamin Johansen, Michael K. Petersen, Jan Larsen, Niels H.
personalized hearing aids, which adapt to the needs of users in real- Pontoppidan, and Jakob E. Larsen. 2018. Learning Preferences and Soundscapes
world listening scenarios. This study addressed the main challenges for Augmented Hearing. In IUI Workshops.
posed when rethinking hearing aids as recommender systems. It [13] Yuki Koyama, Issei Sato, Daisuke Sakamoto, and Takeo Igarashi. 2017. Sequential
Line Search for Efficient Visual Design Optimization by Crowds. ACM Trans.
investigated how to effectively optimize the device settings by Graph. 36, 4, Article 48 (July 2017), 11 pages. https://doi.org/10.1145/3072959.
gathering user preferences in real-world environments. A complex 3073598
audiological space was simplified by decoupling three audiological [14] Nicolas Le Goff. 2015. Amplifying Soft Sounds - a Personal Matter. Technical
Report. Oticon A/S, Smørum, Denmark.
parameters and allowing the participants to manipulate one param- [15] Nicolas Le Goff, Jesper Jensen, Michael S. Pedersen, and Susanna L. Callaway.
eter at a time, comparing four discrete levels. The participants se- 2016. An Introduction to OpenSound NavigatorTM . Technical Report. Oticon A/S,
quentially optimized the three audiological parameters, which were Smørum, Denmark.
[16] Nicholas A. Lesica. 2018. Why Do Hearing Aids Fail to Restore Normal Auditory
subsequently combined into a personalized device configuration. Perception? Trends in Neurosciences 41, 4 (April 2018), 174–185. https://doi.org/
This configuration was blindly compared against a configuration 10.1016/j.tins.2018.01.008
[17] Jeremy Marozeau and Mary Florentine. 2007. Loudness Growth in Individual
personalized in a standard clinical workflow based on questions Listeners with Hearing Losses: A Review. The Journal of the Acoustical Society of
and pre-recorded sound samples, and six out of seven participants America 122, 3 (2007), EL81–EL87. https://doi.org/10.1121/1.2761924
preferred the device settings learned in real-world listening en- [18] Elaine H. N. Ng, Mary Rudner, Thomas Lunner, Michael Syskind Pedersen, and
Jerker Rönnberg. 2013. Effects of Noise and Working Memory Capacity on
vironments. Thus, the approach seemed to effectively gather the Memory Processing of Speech for Hearing-aid Users. International Journal of
main individual audiological preferences. The parameters resulted Audiology 52, 7 (2013), 433–441. https://doi.org/10.3109/14992027.2013.776181
16
Rethinking Hearing Aids as Recommender Systems HealthRecSys’19, September 20, 2019, Copenhagen, Denmark
[19] World Health Organization. 2013. Multi-Country Assessment of National Capacity [22] Jonathan E. Peele and Arthur Wingfield. 2016. The Neural Consequences of
to Provide Hearing Care. Geneva, Switzerland. Age Related Hearing Loss. Trends in Neurosciences 39, 7 (2016), 486–497. https:
[20] World Health Organization. 2019. Deafness and Hearing Loss. Retrieved June //doi.org/10.1016/j.tins.2016.05.001
30, 2019 from https://www.who.int/news-room/fact-sheets/detail/deafness-and- [23] Dorothea Wendt, Renskje K. Hietkamp, and Thomas Lunner. 2017. Impact of
hearing-loss Noise and Noise Reduction on Processing Effort. Ear and Hearing 38, 6 (2017),
[21] Antti Oulasvirta, Per O. Kristensson, Xiaojun Bi, and Andrew Howes. 2018. 690–700. https://doi.org/10.1097/aud.0000000000000454
Computational Interaction. Oxford University Press.
17