Abstract: Conventional audiograms must proceed to completion in order to achieve validated results. This process wastes time when only a simple question is being asked, such as whether two audiograms differ from one another. A machine learning model intended to ask this question directly was designed using a large database of over 2 million audiograms and evaluated for its efficiency. On average, 6 tones were required to make this determination with high confidence. Thus, a rigorous diagnostic test of hearing loss can be conducted with a tone count comparable to screening tests.
Summary:Objective To determine the feasibility and efficiency of a machine learning model for actively deciding whether two audiograms reflect the same or a different amount of hearing loss.
Rationale Most diagnostic tests proceed in serial fashion, with data collection continuing until terminated according to validated standards before any inference is drawn. Inefficient and/or underpowered procedures result. An example of this phenomenon is surveillance monitoring of audiometric thresholds to detect progressive or occupational hearing loss. Only after each identical repeat audiogram is complete can decisions be made about whether new hearing losses have occurred. An improved process would allow data to be collected that specifically supports the desired clinical decision: difference detection.
Methods A probabilistic machine learning model of the audiogram was developed. A large database of over 2 million audiograms was used to quantify seven standard hearing loss phenotypes using this model. All pairs of these phenotypes were then used to generate simulated new data in order to compare a reference audiogram model with a new audiogram model. Optimal tone deliveries were computed in real time to rapidly answer the question of whether the two audiogram phenotypes were the same or different. The mean number of tones required to reach a decision of “same” or “different” was quantified for all phenotype combinations.
Results All conditions of “same” and “different” were accurately determined (i.e., accuracy = 100%). The mean number of actively acquired tones necessary to state with high confidence that two audiogram phenotypes were the same (Bayes factor = 0.01) or different (Bayes factor = 100) was 6 tones. In general, more similar phenotypes required more tones to reach this conclusion than less similar phenotypes.
Conclusion A machine learning method to deliver tones most informative for determining whether two simulated ears reflect the same or different hearing was successfully demonstrated. The small number of tones needed for this determination represents a compelling reason for machine learning audiometry to become more widely practiced. Repeat testing can make use of the patient’s previous audiogram as the reference to detect a change, while new testing can make use of the normal audiogram as the reference in order to diagnose an absolute hearing loss. Therefore, the active difference audiogram can be effectively and efficiently applied under multiple clinical scenarios.
Learning Objectives:
Identify and justify one advantage of machine learning algorithms over conventional procedures for hearing loss assessment.
