Personalizing over-the-counter hearing aids using pairwise comparisons.

Over-the-counter hearing aids enable more affordable and accessible hearing health care by shifting the burden of configuring the device from trained audiologists to end-users. A critical challenge is to provide users with an easy-to-use method for personalizing the many parameters which control sound amplification based on their preferences. This paper presents a novel approach to fitting hearing aids that provides a higher degree of personalization than existing methods by using user feedback more efficiently. Our approach divides the fitting problem into two parts. First, we discretize an initial 24-dimensional space of possible configurations into a small number of presets. Presets are constructed to ensure that they can meet the hearing needs of a large fraction of Americans with mild-to-moderate hearing loss. Then, an online agent learns the best preset by asking a sequence of pairwise comparisons. This learning problem is an instance of the multi-armed bandit problem. We performed a 35-user study to understand the factors that affect user preferences and evaluate the efficacy of multi-armed bandit algorithms. Most notably, we identified a new relationship between a user's preference and presets: a user's preference can be represented as one or more preference points in the initial configuration space with stronger preferences expressed for nearby presets (as measured by the Euclidean distance). Based on this observation, we have developed a Two-Phase Personalizing algorithm that significantly reduces the number of comparisons required to identify a user's preferred preset. Simulation results indicate that the proposed algorithm can find the best configuration with a median of 25 comparisons, reducing by half the comparisons required by the best baseline. These results indicate that it is feasible to configure over-the-counter hearing aids using a small number of pairwise comparisons without the help of professionals.