Bayesian Models of Speech Understanding for a Clinical Population

Authors: Malcolm Slaney1, Matthew Fitzgerald1

1Stanford University

Background: We use Bayesian modeling to connect audiology tests such as speech in noise (SPIN) and not only predict missing data, but also reason about the quality of our models. Modern auditory models describe everything from the mechanics to cortical processing, and include thousands of parameters. We can’t perform every audiological or psychoacoustic test on a patient. Yet we want the data from these tests so we can tune the large number of parameters in these detailed auditory models. A Bayesian model allows us to statistically model the connections between different kinds of data. The output of the Bayesian process is a directed graph of probability distributions that allow us to reason about the distribution of missing data. Most importantly the final distributions are not Gaussian, so we use new efficient sampling methods to estimate these unknown distributions. Here we show results from a preliminary Bayesian model of SPIN performance using clinical data.

Methods: Data were taken from an audiologic database with 90,000 patients. In addition to the normal demographic and threshold measurements we have supra-threshold measurements such as word-recognition in quiet (WRQ) and SPIN. We then built Bayesian models of SPIN performance.

Results: Here we present two preliminary results. We first show a model that combines demographic and threshold data to predict SPIN data as a function of age, even for patients for which we have no test data. We then show a more sophisticated model that assumes that the hearing threshold limits hearing before any cognitive effects manifest themselves. The cognitive effects are modeled with a skewed-normal distribution so cognitive degradation is a number that declines from 1.

Conclusions: While preliminary, these efforts demonstrate the feasibility of Bayesian modeling. Moreover, they reflect a vital step towards parameterization of a full auditory model from only a subset of audiologic and psychoacoustic tests