A flexible data-driven audiological patient stratification method for deriving auditory profiles.

Area of Science:

• Audiology
• Data Science
• Machine Learning

Background:

• Characterizing hearing deficits requires more than audiograms, encompassing various auditory functions.
• Extensive audiological test batteries are often impractical in clinical settings due to time constraints.
• Current clinical practices use limited, variable audiological measures for patient assessment.

Purpose of the Study:

• To develop a flexible, data-driven approach for stratifying patients into distinct auditory profiles.
• To create clinically applicable auditory profiles using a reduced set of commonly available audiological measures.
• To identify an optimal classification model for accurate patient stratification.

Main Methods:

• A prototypical database (N=595) with audiogram data, loudness scaling, speech tests, and anamnesis was used.
• Model-based clustering was employed to define patient groups (auditory profiles).
• Random forest classification models were built and optimized using various parameterizations (binarization, cross-validation, evaluation metrics).

Main Results:

• A set of 13 audiologically plausible auditory profiles was generated, including normal hearing and varying degrees of impairment.
• An optimal random forest model was identified using a hybrid binarization strategy, 10-fold cross-validation, and the kappa metric.
• The optimal model achieved high classification performance: mean precision of 0.9 and mean sensitivity of 0.84 for 12 of 13 profiles.

Conclusions:

• The proposed data-driven approach generates interpretable and clinically applicable auditory profiles for efficient hearing deficit characterization.
• This method supports patient stratification based on audiological data, enhancing clinical routine.
• The approach is flexible for application to diverse audiological datasets and can be expanded to include aided measurements and fitting parameters.