(PP307) Evaluation of Inverse Reinforcement Learning for Achieving Personalized Amplification

The objective of this study was to evaluate a human-in-the-loop inverse reinforcement learning (IRL) framework for developing personalized amplification based on listener preferences. Studies have shown that as many as 50% of hearing aid users prefer settings that differ from prescribed values. Currently, the use of standardized prescriptive fitting formulas, such as the Desired Sensation Level version 5 (DSLv5), is considered best practice. In this study we applied IRL to modify the default compression ratio of DSLv5 across frequencies for 10 hearing impaired users. Our results showed that IRL values were not only preferred but produced higher word recognition scores.  

Summary: 

Objective and rationale: The overall goal of this study was to evaluate a human-in-the-loop inverse reinforcement learning (IRL) framework aimed at achieving personalized amplification. Studies have shown that as approximately half of hearing aid users prefer amplification values that differ from those prescribed. The current standard of care uses a prescriptive approach to apply amplification values such as the Desired Sensation Level version 5 (DSLv5) with little to no user input. This prescriptive approach is based on the average needs of individuals who are hearing impaired. Among these average-based prescriptive values, a critical component is the compression ratio which is used to apply non-linear gain across the frequency range so that sounds that are below threshold are audible and sounds that exceed threshold have a slower amplification ramp to balance both audibility and comfort. In order to personalize DSLv5 prescriptive values, we used an IRL framework to modify compression ratios based on user feedback and evaluated whether these values were preferred over the standard DSLv5 values and whether the preferred values had an effect on word recognition scores.    

Design: This study recruited 10 adult participants. Audiometric pure tone thresholds and DSLv5 amplification targets were used to determine the initial compression values.  Following audiological evaluation, participants completed a computerized listening task through hearing aids connected to a computer via Bluetooth (gain and hearing aid features were disabled). Participants listened to 7 sets of 30 different spoken sentences. Each sentence played twice, first using the initial compression ratios and then again using a second set of compression ratios determined by IRL. The order of the standard and IRL sentences were randomized. For each sentence pair, the participant was prompted to select which presentation they preferred. These data were used to train the algorithm. Based on the participant’s response, the software adjusted the compression ratios and a new sentence was played. The participants completed this task about 200 times or trials and the user’s preferred compression ratios in five frequency ranges were determined by learning the preferred compression ratios of the trials via IRL. Once the preferred compression ratios were identified, the user’s preferences were compared to the standard DSL v5 compression ratios. Finally, the participants completed a word recognition task using the custom compression ratios and again using the DSLv5 compression ratios to assess potential differences in word recognition performance (NU-6) between the two approaches.

Results: All study participants preferred IRL-derived compression ratios when compared to the default DSLv5 prescribed values. In addition, we found that when using the IRL-derived values, word recognition scores were the same or better than scores obtained from the standard DSL v5 values.

Conclusions:  The results presented here suggest that IRL-derived personalized compression ratios were not only preferred but maintained or improved word recognition scores. The use of this technology could improve hearing aid outcomes and satisfaction by incorporating individualized preferences without compromising audibility