Explainable machine learning reveals the relationship between hearing thresholds and speech-in-noise recognition in...

Area of Science:

• Audiology
• Machine Learning
• Signal Processing

Background:

• Some individuals with normal audiograms experience difficulties understanding speech in noisy environments.
• This phenomenon suggests that standard hearing tests may not fully capture all aspects of auditory function relevant to real-world listening.

Purpose of the Study:

• To evaluate the predictive power of hearing thresholds on speech-in-noise recognition in individuals with clinically normal audiograms.
• To compare the performance of various machine learning models (GAM, ANN, DNN, RF, XGBoost) against a standard statistical model.
• To identify the relative importance of specific audiometric frequencies and demographic factors in predicting speech recognition thresholds (SRTs).

Main Methods:

• Applied machine learning models including generalized additive model (GAM), artificial neural network (ANN), deep neural network (DNN), random forest (RF), and eXtreme gradient boosting (XGBoost).
• Utilized archival data from 764 participants (1528 ears) with normal audiograms, including hearing thresholds from 0.25 to 16 kHz and SRTs.
• Employed SHapley Additive exPlanations (SHAP) to determine the contribution of different variables to SRT prediction.

Main Results:

• XGBoost demonstrated the best performance among machine learning models, with a mean absolute error (MAE) of 1.62 dB.
• GAM achieved a comparable MAE of 1.61 dB, while ANN and RF showed similar results (MAE ≈ 1.67-1.68 dB).
• SHAP analysis identified age and hearing thresholds at extended high frequencies (16 kHz, 12.5 kHz) as significant predictors of SRT.

Conclusions:

• Machine learning models, especially XGBoost, can effectively predict speech-in-noise recognition abilities in individuals with normal audiograms.
• Hearing sensitivity in the extended high-frequency range (above 8 kHz) plays a crucial role in speech perception in noise, even with a normal conventional audiogram.
• These findings underscore the need to consider extended high-frequency hearing and demographic factors for a comprehensive assessment of auditory function.