Conductive hearing loss affects all ages and represents over 50% of hearing impairments, but unlike
sensorineural loss, the potential for treatment is high. Conductive loss stems from a diverse set of possible
pathologies, such as ossicular fixation, ossicular disarticulation, or superior-canal dehiscence, each of which
requires a different treatment. Moreover, these distinct pathologies can result from similar physical traumas and
exhibit similar symptoms, which means that in most cases x-ray-based imaging and exploratory surgeries are
used to confirm a suspected pathology. Because of the high cost, risk to the patient, and subjectivity of existing
diagnostic options, an inexpensive, noninvasive measure would be valuable to assess the middle-ear (ME)
status, to reduce uncertainties about the diagnosis prior to surgery, and to monitor outcomes postoperatively.
Wideband tympanometry (WBT), which uses an ear-canal probe to quickly measure the frequency-varying
admittance/impedance of the ME across a range of negative and positive static pressures, could become a cost-
effective tool for noninvasively diagnosing ME pathologies. However, the task of mining complex WBT datasets
for reliable indicators of ME pathologies has proven challenging. Machine learning (ML), with its powerful pattern-
recognition and classification capabilities, may provide a reliable methodology for doing this. However, only very
limited attempts have been made thus far to incorporate ML into ME assessments, mainly due to the lack of
large-enough WBT datasets of confirmed pathologies that are usually required to train ML algorithms. We
propose to train an inference neural network (NN) to perform fast and accurate objective interpretations of WBT
data. To account for the lack of sufficient pathology-identified training data, we propose using synthetic WBT
responses from anatomically realistic finite-element (FE) models of the human ear with verified mechanistic
behavior. Randomly varying the material properties and geometric parameters of the models within normal and
beyond-normal ranges will mimic normal and pathological conditions while accounting for inter-subject variability,
age-related changes to the ME structures, and measurement noise. The inference NN will be trained on this
population of model parameters and responses to produce a probability distribution for each parameter value
whenever it is presented with a new WBT response. Since each model parameter maps to a specific
physiological characteristic of the ME, the predicted parameter values can indicate whether a response exhibits
normal or pathological characteristics. Next, the NN knowledge will be expanded by applying transfer learning
to the limited available clinical WBT data of confirmed pathological cases, along with additional noninvasive
clinical data such as audiograms and air–bone gap measurements. The outcome of the project will be a trained
inference NN for noninvasive objective assessments of the likelihood that a given ear has one (or more) of
various conductive pathologies. Its use could reduce the need for or avoid unnecessary exploratory surgery,
improve the specificity of preoperative preparations, and provide a low-cost means of postoperative monitoring.
