Bayesian Pure-Tone Audiometry Through Active Learning Under Informed Priors

Marco Cox, Bert de VriesFrontiers Digit. Health 2021 · DOI

TL;DR

This paper introduces a more sophisticated Bayesian approach to pure-tone audiometry (PTA), the standard method for assessing hearing. By using a mixture of Gaussian Processes (GPs) that can be conditioned on subject-specific information like age and gender, they achieve significantly more efficient and accurate hearing threshold estimation compared to traditional single-GP models.

What Problem Does This Paper Address?

Pure-tone audiometry is crucial for diagnosing and quantifying hearing loss. The process involves presenting tones of varying frequencies and intensities and recording whether the subject hears them. The goal is to estimate the hearing threshold (the minimum intensity audible) at different frequencies. Traditional methods often rely on heuristic protocols, which can be time-consuming and may not be optimal for every individual. While Bayesian methods and Gaussian Processes (GPs) have shown promise in making PTA more efficient by actively selecting the most informative stimuli, the accuracy of these methods is heavily dependent on the quality of the underlying probabilistic model. This paper addresses the limitation of single-GP models by proposing a more flexible and data-driven approach to model hearing thresholds, aiming to improve both the accuracy and efficiency of PTA.

Methods

The core of this work is a Bayesian probabilistic model for hearing threshold estimation. Instead of a single Gaussian Process (GP) to model hearing thresholds across all frequencies, the authors propose a Gaussian Process Mixture Model (GPMM). The key innovations are:

  • Mixture of GPs: The model comprises multiple independent GPs, each potentially representing a different "type" of hearing threshold (e.g., different slopes or shapes). This allows for a more nuanced and accurate representation of the diverse range of hearing thresholds found in the population.
  • Conditioning on Side-Information: The mixing weights of these GPs are not fixed but can be conditioned on subject-specific information like age and gender. This allows the model to adapt its prior beliefs about the hearing threshold based on demographic factors known to influence hearing.
  • Active Learning: The model is integrated into an active learning framework. This means that at each step of the audiometry test, the system intelligently selects the next tone (frequency and intensity) that is expected to provide the most information about the subject's hearing threshold, thereby minimizing the total number of tones needed. This selection is based on maximizing expected information gain.
  • Bayesian Inference: After each response, the model is updated using Bayesian inference, refining the estimate of the hearing threshold and its uncertainty.
  • Model Learning: The parameters of the GPMM (including the individual GP kernels and the mixing weight functions) are learned from a large dataset of annotated audiograms using maximum likelihood estimation. This data-driven optimization tailors the model to the specific population from which the data was drawn.

Datasets

The paper utilizes a proprietary dataset containing approximately 176,474 annotated audiograms from the Nordic countries. This dataset includes the hearing thresholds at various standard frequencies, along with the age and gender of the subjects. This dataset is used for learning the parameters of the GPMM. For evaluation, the data is split into an 80% training set and a 20% test set.

Key Insights and Ideas

  • Heterogeneity of Hearing Thresholds: The paper's central idea is that hearing thresholds are not monolithic. Different patterns exist, and a mixture model can capture this heterogeneity much better than a single model.
  • Demographics Matter: Age and gender are significant factors influencing hearing thresholds. Conditioning the model on these attributes allows for a more personalized and accurate prior, leading to better predictions.
  • Data-Driven Optimization is Crucial: Simply having a complex model isn't enough. Optimizing the model's parameters on a large, representative dataset is essential for it to effectively capture the statistical properties of hearing thresholds.
  • Active Learning Efficiency: The combination of a good probabilistic model and an active learning strategy dramatically reduces the number of stimuli required to achieve a reliable hearing threshold estimate.
  • Convergence of Mixing Weights: During simulations, the posterior mixing weights of the GPMM tend to converge to either 0 or 1 as more responses are collected. This suggests the model effectively identifies the most relevant GP component for a given subject and then refines the estimate using that component.

Key Results

  • Improved Predictive Accuracy: The optimized GPMM significantly outperforms an optimized single-GP model in terms of predictive accuracy, as measured by the average posterior log-likelihood on a test set. The improvement is more pronounced with a larger number of mixture components (up to around seven).
  • Faster Convergence: In simulations, the 8-component GPMM (with or without age/gender conditioning) achieved a substantially faster reduction in average absolute HT estimation error compared to the single-component model.
  • Benefit of Conditioning: Conditioning the GPMM on age and/or gender further improves predictive accuracy and reduces estimation error, especially when the number of mixture components is sufficient. Age conditioning showed a stronger impact than gender conditioning.
  • Efficiency Gains: The simulations suggest that the proposed Bayesian PTA method, particularly with the GPMM, leads to significant increases in efficiency, requiring fewer responses to reach a desired accuracy level.

Limitations and Future Work

  • Computational Complexity: The GPMM increases computational complexity linearly with the number of mixture components. While manageable, it's a trade-off for improved performance.
  • User Response Model: The authors suggest that the user response model (how a subject's response is generated given their HT) could be made more sophisticated by incorporating frequency dependence.
  • Robustness to Corrupted Responses: The current model is not explicitly designed to handle corrupted or erroneous responses. Future work could involve developing methods to detect and exclude such responses or incorporate a data corruption aspect into the model, though this might slow down convergence.
  • Inter-Ear Correlation: The paper notes a statistically significant correlation between the hearing thresholds of the two ears, which could be exploited in future models to improve efficiency for testing the second ear.

Summary updated

BibTeX
@article{cox2021bayesian,
  author = {Marco Cox and Bert de Vries},
  title = {{Bayesian Pure-Tone Audiometry Through Active Learning Under Informed Priors}},
  journal = {Frontiers Digit. Health},
  year = {2021},
  doi = {10.3389/fdgth.2021.723348},
}
Figure