Dissertation Abstract: Hearing Aid Fitting with Genetic Algorithms

by Eric Alan Durant

Chair: Gregory H. Wakefield

Hearing aids are controlled by numerous parameters, presenting the audiologist with a difficult fitting task. While many of these parameters can be set near their optimal values using prescriptive formulas based on simple measurements of the patient’s hearing loss, others are a matter of comfort and other qualities that the patient must evaluate subjectively. Thus, there is much interest in including the patient’s preferences as a key element of the fitting process.

To accomplish this, various researchers have looked to mathematical optimization theory. As the number and complexity of hearing aid parameters increase, the shortcomings of these methods become burdensome. To overcome these shortcomings, we investigate the genetic algorithm (GA). The GA is a search procedure that borrows many concepts from biology, including natural selection and genetic crossover and mutation. The GA maintains a population of solutions (hearing aid parameter sets) and repeatedly replaces the least fit solutions with the offspring of better performing solutions. In this thesis, methods are developed for applying the GA to hearing aid fitting. These include an efficient procedure to determine the relative quality of solutions in the population by repeatedly asking the patient to select the better of two alternatives.

To evaluate our approach, we conducted two experiments with eight normal hearing and eight hearing-impaired subjects. In the first, three parameters were varied to control the cancellation of feedback, a common problem in which objects near the aid and certain motions cause it to squeal. In the second, six parameters were varied to fit three-band dynamic range expansion, which controls tradeoffs between suppressing unwanted background sounds and sufficiently amplifying desired sounds such as speech.

We found that the GA worked very well for fitting the feedback cancellation system, using both objective and subjective measures. In addition, we learned that patients have greatly differing preferences for feedback cancellation parameters and that these preferences do not change much when subjects are retested. The results for the expansion system were also positive, but highlighted some problems and suggested changes that might be made to improve performance when fitting more complicated parameter sets.