Big Data Analysis on Coupling Selection and Model Development of a Data Driven Dome Proposer

Authors: Richard Rau1, Stefan Pislak1, Volker Kuehnel1

1Sonova AG

Background: Selecting the optimal acoustic coupling is essential for user acceptance of hearing devices, involving nuanced trade-offs among occlusion effect, preferences for direct versus amplified sound, feedback risk, and performance of beamforming and noise reduction. These factors are deeply interconnected, complicating the creation of deterministic rule-based algorithms. To address this challenge, we propose a data-driven approach that employs a neural network model trained on historical fitting data to predict the best acoustic couplings. This study also explores parameters previously not included in prior models.

Methods: We processed and analyzed 50,000 logged fittings from the “Phonak Target” software after cleansing the data. We identified critical parameters influencing coupling selection and developed a neural network model accordingly. Each data point was scored for expected user satisfaction and intelligibility, derived from daily wear time and the aided Speech Intelligibility Index (Fig. 1a). To reduce bias, scores were normalized across similar hearing loss profiles, and the dataset was augmented to favor samples with higher scores.

Results: Our research shows that user experience and contralateral hearing loss, factors previously overlooked, influence coupling preferences. The newly developed model demonstrated a notable 5 percentage point increase in acceptance rate on a holdout test set compared to the prior model. Notably, it more accurately proposed Power and Vented Domes, aligning better with actual user choices (Fig. 1b&c).

Conclusions: The neural network model enhances the precision of acoustic coupling recommendations, providing a robust foundation for future advancements in hearing device fittings. Further investigations into additional user-specific parameters could improve the model’s effectiveness and user satisfaction even more.