Data-driven segmentation of audiometric phenotypes across a large clinical cohort

Area of Science:

• Audiology
• Data Science
• Biostatistics

Background:

• Pure tone audiograms are standard for assessing hearing loss.
• Current categorization methods are limited, failing to classify nearly half of patient records.
• Existing typologies assume specific relationships and etiologies, overlooking data variability.

Purpose of the Study:

• To develop a more comprehensive classification system for hearing loss profiles.
• To analyze a large dataset of audiograms using an unsupervised machine learning approach.
• To identify novel audiogram types and their correlation with demographic factors.

Main Methods:

• Analysis of 116,400 patient audiograms over 24 years.
• Application of Gaussian Mixture Models (GMM) for unsupervised audiogram segmentation.
• Validation using 15,380 audiograms from the National Health and Nutrition Examination Survey (NHANES).

Main Results:

• Standard categorization left 46% of patient records unclassified.
• GMM identified ten distinct audiogram types, including high-frequency, flat, mixed, and notched profiles.
• Identified audiogram types showed predictable relationships with patient age and sex.
• NHANES data yielded six similar types, lacking extreme configurations found in the clinical cohort.

Conclusions:

• An objective, data-driven GMM approach provides a more nuanced and comprehensive classification of hearing loss than traditional methods.
• The identified audiogram types are consistent across independent clinical and population-based datasets.
• This probabilistic model accounts for audiogram variability, offering a richer understanding of hearing loss heterogeneity.