Authors: Mériem Jaïdane¹, Sadok Gaied², Rim Amara³
¹Ecole Nationale d’Ingénieurs de Tunis
²INSAT & dB.Sense
³L3S-ENIT & INSAT
There is a growing interest among healthcare professionals (ENTs and audioprosthesists) in computational audiology approaches. While faster acquisition of continuous audiograms is appreciated, probabilistic representations of a patient’s audiogram are not very common.
We hypothesize that adopting probabilistic audiometry involves -for example- exploring standard protocols, such as the modified Hughston-Westlake (HW) protocol, which tests a finite set of predefined frequencies [1]. We show that this protocol already enables estimation of a continuous audiogram with induced accuracy by exploiting unused intermediate data.
Our approach includes two main steps:
Auditory response simulator using the HW protocol: It generates synthetic patient data by simulating a test sequence. Key hyperparameters include: dB level of initial familiarization tone, descent/ascent steps, number of repeated ascending responses needed to identify threshold, and the psychometric function representing patient hearing characteristics. This simulation is applied across the set of frequencies tested.
Gaussian process inference: it is a component of the Bayesian Active Learning method [2] used in machine learning for multidimensional perceptual/latent functions (audiogram determination) from binary responses data. Assuming an underlying Gaussian process, we estimate the audiogram non-parametrically, using the measurement points generated by the new proposed HW simulator.
Simulations (e.g fig. 1 and fig. 2) illustrate that the probabilistic model of the audiometric curve, built from typically unused HW protocol data, closely fits the original threshold points. The resulting continuous curves are both familiar and informative, providing uncertainty estimates in untested regions—thanks to the continuity constraints imposed by the Gaussian process covariance matrix.
