Active transfer learning for audiogram estimation

Abstract

Computational audiology (CA) has grown over the last few years with the improvement of computing power and the growth of machine learning (ML) models. There are today several audiogram databases which have been used to improve the accuracy of CA models as well as reduce testing time and diagnostic complexity. However, these CA models have mainly been trained on single populations. This study integrated contextual and prior knowledge from audiogram databases of multiple populations as informative priors to estimate audiograms more precisely using two mechanisms: (1) a mapping function drawn from feature-based homogeneous Transfer Learning (TL) also known as Domain Adaptation (DA) and (2) Active Learning (Uncertainty Sampling) using a stream-based query mechanism. Simulations of the Active Transfer Learning (ATL) model were tested against a traditional adaptive staircase method akin to the Hughson-Westlake (HW) method for the left ear at frequencies ω=0.25,0.5,1,2,4,8 kHz, resulting in accuracy and reliability improvements. ATL improved HW tests from a mean of 41.3 sound stimuli presentations and reliability of ±9.02 dB down to 25.3±1.04 dB. Integrating multiple databases also resulted in classifying the audiograms into 18 phenotypes, which means that with increasing data-driven CA, higher precision is achievable, and a possible re-conceptualisation of the notion of phenotype classifications might be required. The study contributes to CA in identifying an ATL mechanism to leverage existing audiogram databases and CA models across different population groups. Further studies can be done for other psychophysical phenomena using ATL.