Improving accessibility, accuracy, and efficiency of auditory evoked response testing using a telehealth and artificial intelligence approach

Abstract Approximately 2 to 3 out of every 1000 newborns in the US experience detectable hearing loss in at least one ear. For young patients from whom behavioral hearing thresholds cannot be obtained, the neurophysiological potentials evoked in response to acoustic stimuli, such as the auditory brainstem response (ABR) measured using scalp electrodes, are often used for hearing diagnosis. Because ABR testing requires specialized equipment and clinical expertise from experienced pediatric audiologists, there is a general lack of access to diagnostic ABR testing, especially in remote and rural communities. The overall goal of this project is to develop an efficient automated ABR testing system incorporating artificial intelligence (AI) to optimize testing for telemedicine applications in hearing screening and diagnostic applications. An AI-based algorithm that adaptively select the suitable test stimuli during ABR testing, developed at the University of Washington, will be integrated with a cost-effective hardware platform developed at Intelligent Hearing Systems Corporation for conducting telehealth-enabled hearing diagnosis. Additionally, the AI-based algorithm will be further improved using a database of hundreds to thousands of clinical ABR test sessions obtained from Seattle Children’s Hospital. Two separate data-driven procedures will be conducted. First, the database will be used to improve the estimation of ABR thresholds, which is the lowest sound level to evoke an observable ABR response and the main indicator for the degree of hearing loss. It is anticipated that the data-driven procedure in ABR threshold estimation will improve the performance of the AI-based algorithm, compared to the prototype algorithm that does not incorporate prior knowledge on ABR data when determining ABR thresholds. Second, a deep-learning model of expert clinicians’ decision-making process will be constructed. The model will take ABR waveform features collected for preceding test stimuli as its input and predict expert clinicians’ choice for the next test stimulus (for example, to increase the test level, to decrease the test level, or to repeat the current level). The model will be implemented with AI-based algorithm to narrow the range for stimulus optimization, hence allowing the AI-based algorithm to make stimulus-selection choice similar to that of expert pediatric audiologists. It is anticipated that the model of clinicians’ behavior during ABR testing will improve the efficiency and robustness of the AI-based algorithm. At the completion of the proposed research, a novel ABR testing system that enables asynchronous tele-diagnostic ABR testing without real-time supervision from an experienced audiologist will be fully developed and ready for clinical evaluations in the field. Such a system would be more reliable, efficient, and cost-effective than the synchronous telehealth approach currently available for diagnostic ABR testing.