Personalizing over-the-counter hearing aids using pairwise comparisons

Abstract

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.

Keywords: Active learning; Hearing aids; Pairwise comparisons; Personalization.

Fig. 1.

Fig. 1(a) shows our approach in two steps: (1) discretize the configuration space to identify 15 representative presets and (2) a personalization algorithm is used to select the most preferred preset based on user feedback. Fig. 1(b) shows block diagram of a learning algorithm.

Fig. 2.

The 8 REARs at 65 dB SPL of the 934 configurations are mapped into a two-dimensional space using PCA and plotted in Fig. 2(a) as gray dots. The blue contours depict the density of REAR configurations in two-dimensions. Fig. 2(b) plots the coverage of an increasing number of presets and shows that 15 presets can cover 95% of the population with mild-to-moderate hearing loss. The selected 15 presets are shown as red dots in Fig. 2(a) and the REARs of some presets are plotted in Fig. 2(c).

Fig. 3.

User interface used to obtain subjects’ feedback.

Fig. 4.

In a pairwise comparison a user may express an equal, weak, or strong preference for one of the two presets. The prevalence of equal, weak, and strong preferences varies significantly across subjects and preset pairs.

Fig. 5.

Distribution of Borda scores for each subject.

Fig. 6.

Borda scores of representative subjects are shown in Fig. 6(a). Fig. 6(b) plots the CDF of the best Borda scores for all subjects. Fig. 6(c) plots the number of presets (excluding the best preset) within 0.05 and 0.1 thresholds.

Fig. 7.

Fig. 7(a) shows the proportion of tuples for which total order, triangle inequality and stochastic transitivity assumptions hold. Fig. 7(b) shows the average of the REARs at frequencies 0.5, 1, 2, 4 kHz (denoted as 4FA REAR) for preferred presets and NAL-NL2 configuration. Each point represents a subject.

Fig. 8.

Figs. 8(a) and 8(b) show the subject’s most preferred setting as a red triangle. The shading of squares indicate the fraction of runs in which a preset won. Presets that won in no runs are shown as circles. Fig. 8(c) compares the distribution of distances between preset winners and a set of randomly chosen presets.

Fig. 9.

TPP’s pseudocode.

Fig. 10.

Fig. 10(a) plots the presets’ location in two dimensions. The dark red squares indicate the subset considered during the global prediction. The green “x” indicates the location of the predicted best location. Fig. 10(b) plots as dark red squares the presets involved in the local search phase. The winning preset is indicated by blue triangle in both figures.

Fig. 11.

Fig. 11(a) shows distribution of regret with TPP and other baselines. Solid black line indicates the regret threshold of 0.05. Fig. 11(b) shows number of comparisons required for median regret to be below 0.05 threshold for varying number of presets.

Fig. 12.

Distribution of regret for each subject after 40 pairwise comparisons with TPP and SAVAGE. Solid black line indicates the regret threshold of 0.05.