Bayesian Pure-Tone Audiometry Through Active Learning Under Informed Priors.

Area of Science:

• Audiology
• Machine Learning
• Statistical Modeling

Background:

• Pure-tone audiometry is fundamental for diagnosing and quantifying hearing loss.
• Gaussian process (GP) models have shown promise in probabilistic audiometry.
• Improving the underlying model is key to enhancing audiometry method performance.

Purpose of the Study:

• To enhance the efficiency of GP-based audiometry procedures.
• To introduce Gaussian process mixture models for improved hearing threshold estimation.
• To condition mixture models on subject-specific side-information like age and gender.

Main Methods:

• Utilized Gaussian process (GP) mixture models instead of single GPs.
• Conditioned mixing coefficients on side-information (age, gender) to capture correlations.
• Optimized GP mixture models by learning parameters from a large dataset of annotated audiograms.
• Derived optimal tone selection via greedy information gain maximization and hearing threshold estimation through Bayesian inference.

Main Results:

• Optimized GP mixture models demonstrated significantly higher predictive accuracy compared to optimized single-GP models.
• The proposed models were fitted to approximately 176,000 annotated audiograms from Nordic countries.
• Audiometry simulations showed substantial increases in the efficiency of Bayesian audiometry procedures using GP mixture models.

Conclusions:

• Gaussian process mixture models offer a significant advancement over single GP models in audiometry.
• Conditioning on side-information improves the statistical modeling of hearing thresholds.
• The developed methods enhance the accuracy and efficiency of hearing loss diagnosis and quantification.