A Machine Learning Based Clustering Protocol for Determining Hearing Aid Initial Configurations from Pure-Tone Audiograms

Chelzy Belitz, Hussnain Ali, John H.l. HansenInterspeech 2019

TL;DR

This paper proposes a machine learning approach to streamline hearing aid (HA) fitting by clustering audiograms and identifying a limited set of optimal initial HA configurations. By reducing the number of manual adjustments needed, this method aims to improve HA adoption and user satisfaction.

What Problem Does This Paper Address?

Millions of people experience hearing loss, yet a significant portion of them do not use hearing aids. This low adoption rate is partly attributed to the complex and time-consuming process of fitting hearing aids, which often requires multiple adjustments to achieve optimal performance and comfort for the individual. The authors highlight that barriers such as cost, insurance, and the necessity of repeat visits contribute to this issue. Furthermore, even among those who own hearing aids, many do not use them due to problems with fit, comfort, or difficulty adjusting to the device. This paper aims to tackle this by developing a more efficient initial HA configuration process, potentially leading to better HA retention and accessibility.

Methods

The core of this research lies in a two-step machine learning protocol.

  1. Clustering Comfort Targets: The first step involves unsupervised learning to group similar hearing aid "comfort targets" (representing desired amplification levels across frequencies) into a limited number of clusters. The authors explored three clustering algorithms:
  2. Agglomerative Clustering (Ward linkage): A hierarchical approach that merges clusters based on minimizing variance.
  3. Birch Clustering: An efficient agglomerative algorithm designed for large datasets.
  4. K-Means Clustering: An iterative algorithm that partitions data into a predefined number of clusters by minimizing the distance to cluster centroids.

    The goal here is to identify representative "centroids" for each cluster, which will serve as the preset initial HA configurations.

  5. Audiogram Mapping: The second step uses supervised learning to map pure-tone audiograms to these identified preset configurations. Each audiogram is assigned to the same cluster as its associated comfort target. Various classification methods were tested to predict the most suitable preset configuration for a given audiogram.

    The paper also introduces a visual representation of the data using t-SNE to project high-dimensional audiogram and comfort target data into a 2D space for better understanding of cluster separation.

Datasets

The study utilized a large dataset comprising over 90,000 audiogram-fitting pairs. This data was collected from hearing centers across the USA and includes eleven test frequencies. For their analysis, the researchers focused on the four core frequencies (500, 1000, 2000, and 4000 Hz) as they were most consistently represented in the dataset.

Key Insights and Ideas

  • Clustering as a Shortcut: The fundamental idea is that by grouping similar comfort targets, we can define a small set of "ideal" starting points for HA fittings, significantly reducing the need for extensive manual adjustments.
  • K-Means Outperforms: Among the tested clustering algorithms, K-Means emerged as the most effective in minimizing the average adjustment needed between the cluster centroid and the actual comfort targets.
  • Audiogram Overlap is a Challenge: Visualizations revealed that audiograms and their associated comfort targets can be highly overlapped in the acoustic space, making precise classification difficult. This explains why classification accuracies, while decent, are not perfect.
  • Top-Two Classification is Powerful: Even with overlapping data, classifying an audiogram into its top two most likely preset configurations provides a significant improvement in narrowing down the fitting options, offering a practical benefit.

Key Results

  • Reduced Adjustment: K-Means clustering resulted in the lowest average total adjustment of 4.89 when using cluster centroids as preset configurations.
  • Classification Accuracy: Various classifiers achieved test accuracies around 63-64% for assigning audiograms to a single cluster.
  • Improved Top-Two Accuracy: When considering the top two most likely classifications, the accuracy significantly increased, reaching up to 92.70% for all classifications and 68.00% for "strong" classifications.
  • Strong Classifications: A significant portion of the data (82.68%) could be strongly classified, meaning the likelihood of belonging to one cluster was substantially higher than others.

Limitations and Future Work

  • Limited Number of Presets: The study intentionally limited the number of clusters to four to create a small, manageable set of presets. However, the authors acknowledge that introducing more clusters could further reduce the average total adjustment.
  • Overlapping Data: The highly overlapped nature of the audiogram data presents a challenge for precise classification, leading to misclassifications.
  • Future Work: The authors suggest exploring a wider range of hearing aids and investigating classification methods that are more robust to highly overlapped datasets. They also propose further analysis on how to best utilize the "strong" and "weak" classification outcomes.

Summary updated

BibTeX
@inproceedings{belitz2019machine,
  author = {Chelzy Belitz and Hussnain Ali and John H.L. Hansen},
  title = {{A Machine Learning Based Clustering Protocol for Determining Hearing Aid Initial Configurations from Pure-Tone Audiograms}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure