Active transfer learning for audiogram estimation.

Area of Science:

• Computational audiology
• Machine Learning
• Data Science

Background:

• Computational audiology (CA) models have advanced with increased computing power and machine learning (ML).
• Existing CA models often lack generalization due to training on single population audiogram databases.
• Accurate audiogram estimation is crucial for hearing health diagnostics.

Purpose of the Study:

• To develop a more precise audiogram estimation method by integrating knowledge from multiple population databases.
• To enhance CA models using Transfer Learning (TL) and Active Learning (AL).
• To evaluate the performance of the proposed Active Transfer Learning (ATL) model against traditional methods.

Main Methods:

• Integrated contextual and prior knowledge from multiple audiogram databases.
• Employed feature-based homogeneous Transfer Learning (TL) for Domain Adaptation (DA).
• Utilized Active Learning (AL) with a stream-based query mechanism for uncertainty sampling.

Main Results:

• The ATL model demonstrated improved accuracy and reliability compared to the traditional Hughson-Westlake (HW) method.
• ATL reduced sound stimuli presentations from a mean of 41.3 to achieve better results.
• Integration of multiple databases enabled classification into 18 audiogram phenotypes, suggesting potential re-conceptualization of classifications.

Conclusions:

• The developed ATL mechanism effectively leverages diverse audiogram databases for improved CA.
• This approach enhances audiogram precision and diagnostic efficiency.
• ATL holds promise for application in other psychophysical phenomena beyond audiology.