Contextually Adapting Hearing Aids by Learning User Preferences from Data

Although hearing aid users perceive sound in individual ways, current approaches do not fully exploit the potential for personalization. Providing a comprehensively personalized hearing aid solution is a complex and multidimensional challenge and requires a deep understanding of patients’ preferences and behavior. This thesis leverages real-world data collected through smartphone-connected hearing aids to address two main themes. First, personalizing hearing aid settings requires learning the audiological preferences of users. We adopted a smartphone-based method to make users explore three audiological parameters (Noise Reduction, Brightness, and Soft Gain) and to gather their audiological preferences in real-world listening environments. The collected data was modeled to investigate the feasibility of a context-aware system for providing users with a number of relevant hearing aid settings to choose from. We found that having access to different intervention levels of two audiological parameters (Brightness and Soft Gain) affected listening satisfaction. Moreover, context significantly impacted the perceived usefulness of having access to different intervention levels, as well as the intervention level preferences. Second, offering a comprehensively personalized solution, as well as transferring the learned audiological preferences to new or inactive users, requires learning users’ behavior. Large scale data logged by commercially available products were analyzed to explore patterns of hearing aid use, as well as the provision
and context of use of listening programs. We found that, on average, users used the hearing aids 10 hours/day and we identified three clusters of users, each characterized by a predominant daily pattern of hearing aid use. Moreover, we identified a default listening program, a primary additional program, and two secondary additional programs. We also found that users used the additional listening programs in sound environments different than the sound environment measured when using the default program.
This thesis contributes to the progress towards a data-driven approach to realtime hearing aid personalization by learning users’ preferences and behavior from data. It also demonstrates that smartphone-connected hearing aids can be useful to both perform experimental studies aimed at exploring novel ways of personalizing the device, and observational studies aimed at investigating how users naturally use commercially available devices in real-world settings.