Article
Predicting Standard Audiograms From a Loudness Scaling Test Employing Unsupervised, Supervised, and Explainable Machine Learning Techniques (en)
* Presenting author
Abstract:
Mobile hearing health applications – such as the remote characterization of a hearing loss or the fitting of a hearing device – can measure the audiogram only with a large uncertainty due to using uncalibrated devices and uncontrolled acoustic environments. Categorical loudness scaling (CLS) and other supra-threshold tests, on the other hand, can be performed with a higher precision. We therefore evaluate the potential of using CLS instead of a remote audiogram for a precise characterization of hearing deficits and the audiogram-based fitting of hearing devices. An automatic machine learning (ML) based classification system was established on a large auditory database (N = 847) to predict the standard audiogram given 12 supra-threshold parameters from a CLS test. To construct it, 7 supervised multi-class ML classifiers were trained, among which the highest balanced accuracy (0.51 ± 0.07) and weighted F1 score (0.56 ± 0.06) was the heuristic logistics regression algorithm. Three post-hoc explainable ML models were applied to examine the importance of 12 features, revealing that the CLS-estimated hearing threshold level played the most important role while the high-level slope was of minor importance. The implications of this approach for mobile hearing health applications will be discussed.