Machine Learning in Audiology: Applications and Implications
François Charih · 2019
Recent mobile and automated audiometry technologies have allowed for the democratization of hearing healthcare and enables non-experts to deliver hearing tests.The problem remains that a large number of such users are not trained to interpret audiograms.In this work, we outline the development of an intelligent audiogram classification system.More specifically, we present how a training dataset was collected, the development of the classification system relying on supervised learning, as well as other tools designed for the analysis of audiograms in large databases.Using a dedicated annotation tool developed specifically for this study, the Rapid Audiogram Annotation Environment, we collected hundreds of audiogram annotations from three licensed audiologists.Our analysis demonstrates that inter-rater reliability is substantial or better for classification of hearing loss configuration, symmetry, and severity, in spite of the subjective nature of the classification task.Furthermore, our results suggest that the agreement nonexistent for the identification of audiometric notches or potentially unreliable thresholds.The system proposed here achieves a performance comparable to the state of the art, but is significantly more flexible.Finally, we demonstrate qualitatively that a method based on density estimation with Gaussian mixture models is useful for the detection of potential reliability issues in audiograms.iiiFirst and foremost, I would like to express my gratitude to my research advisor, Prof. James Green, for his guidance, his support during the difficult times, and for his unwavering patience.I thank him for continuously pushing me to take an extra step out of my comfort zone so that I can see a little farther and a little clearer.