Improving the classification of noise haters and distortion haters with mixture latent state-trait autoregressive modelling

Authors: Giulia Angonese1,2, Mareike Buhl3, Jonathan A. Gößwein4, Birger Kollmeier1, Andrea Hildebrandt1

1Carl von Ossietzky Universität Oldenburg
2Cluster of Excellence ‘Hearing4all’
3Center for Research and Innovation in Human Audiology (CERIAH), Institut de l’Audition, Institut Pasteur
4Fraunhofer Institute for Digital Media Technology

Background: Individuals have different preferences for setting hearing aid (HA) algorithms that reduce ambient noise but introduce signal distortion. “Noise haters” prefer greater noise reduction, even at the expense of signal quality. “Distortion haters” accept higher noise levels to avoid signal distortion. So far, these preferences have been considered stable over time and individuals have been classified solely on the basis of their trait scores. However, the question remains as to how stable individual listening preferences are and whether state-related variability needs to be considered as a further criterion for classifying individuals according to their preferences.

Methods: We designed a mobile task to measure noise-distortion preferences over two weeks in an ecological momentary assessment study with N=185 (106 f, Mage=63.1, SDage=6.5) unaided individuals who reported subjective hearing difficulties. Latent State-Trait Autoregressive (LST-AR) modelling was used to evaluate the stability of listening preferences and explore temporal dependencies between consecutive observations. The model has been extended to Mixture LST-AR for data-driven classification of noise and distortion haters, taking into account trait and state components of listening preferences.

Results: Individual listening preferences show considerable state-related variance. Information on state-trait mean and variability was used for data driven classification. In addition to successful identification of noise haters, distortion haters and a third intermediate class based on longitudinal, outside of the lab data, we further differentiated between stable versus more variable individuals on listening preferences across multiple days.

Conclusions: Individualisation of HA fitting could be improved by assessing individual preferences along the noise-distortion trade-off, but we also need to take into account the day-to-day variability of these preferences in some individuals more than in others.