Date of Award
Doctor of Philosophy (PhD)
Hearing loss is a critical public health concern, affecting hundreds millions of people worldwide and dramatically impacting quality of life for affected individuals. While treatment techniques have evolved in recent years, methods for assessing hearing ability have remained relatively unchanged for decades. The standard clinical procedure is the modified Hughson-Westlake procedure, an adaptive pure-tone detection task that is typically performed manually by audiologists, costing millions of collective hours annually among healthcare professionals. In addition to the high burden of labor, the technique provides limited detail about an individual’s hearing ability, estimating only detection thresholds at a handful of pre-defined pure-tone frequencies (a threshold audiogram). An efficient technique that produces a detailed estimate of the audiometric function, including threshold and spread, could allow for better characterization of particular hearing pathologies and provide more diagnostic value. Parametric techniques exist to efficiently estimate multidimensional psychometric functions, but are ill-suited for estimation of audiometric functions because these functions cannot be easily parameterized.
The Gaussian process is a compelling machine learning technique for inference of nonparametric multidimensional functions using binary data. The work described in this thesis utilizes Gaussian process classification to build an automated framework for efficient, high-resolution estimation of the full audiometric function, which we call the machine learning audiogram (MLAG). This Bayesian technique iteratively computes a posterior distribution describing its current belief about detection probability given the current set of observed pure tones and detection responses. The posterior distribution can be used to provide a current point estimate of the psychometric function as well as to select an informative query point for the next stimulus to be provided to the listener. The Gaussian process covariance function encodes correlations between variables, reflecting prior beliefs on the system; MLAG uses a composite linear/squared exponential covariance function that enforces monotonicity with respect to intensity but only smoothness with respect to frequency for the audiometric function.
This framework was initially evaluated in human subjects for threshold audiogram estimation. 2 repetitions of MLAG and 1 repetition of manual clinical audiometry were conducted in each of 21 participants. Results indicated that MLAG both agreed with clinical estimates and exhibited test-retest reliability to within accepted clinical standards, but with significantly fewer tone deliveries required compared to clinical methods while also providing an effectively continuous threshold estimate along frequency. This framework’s ability to evaluate full psychometric functions was then evaluated using simulated experiments. As a feasibility check, performance for estimating unidimensional psychometric functions was assessed and directly compared to inference using standard maximum-likelihood probit regression; results indicated that the two methods exhibited near identical performance for estimating threshold and spread. MLAG was then used to estimate 2-dimensional audiometric functions constructed using existing audiogram phenotypes. Results showed that this framework could estimate both threshold and spread of the full audiometric function with high accuracy and reliability given a sufficient sample count; non-active sampling using the Halton set required between 50-100 queries to reach clinical reliability, while active sampling strategies reduced the required number to around 20-30, with Bayesian active leaning by disagreement exhibiting the best performance of the tested methods. Overall, MLAG’s accuracy, reliability, and high degree of detail make it a promising method for estimation of threshold audiograms and audiometric functions, and the framework’s flexibility enables it to be easily extended to other psychophysical domains.
Dennis L. Barbour
Steven E. Petersen, Baranidharan Raman, Mitchell S. Sommers, Kurt A. Thoroughman,