Deep Learning Models for Predicting Hearing Thresholds Based on Swept-Tone Stimulus-Frequency Otoacoustic Emissions

Area of Science:

• Audiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Hearing threshold prediction is crucial for diagnosing hearing loss.
• Stimulus-frequency otoacoustic emissions (SFOAEs) offer a non-invasive measure of cochlear function.
• Traditional methods for hearing threshold prediction from SFOAEs have limitations.

Purpose of the Study:

• To develop and evaluate deep learning (DL) models for quantitative prediction of hearing thresholds.
• To assess the performance of DL models using SFOAEs evoked by swept tones.
• To compare the efficacy of DL models against traditional machine learning (ML) approaches.

Main Methods:

• Four DL models (CNN, CNN-KNN, CNN-SVM, CNN-RF) were developed to predict hearing thresholds.
• Models utilized SFOAE amplitude spectra and signal-to-noise ratio spectra as input.
• Performance was evaluated using mean absolute error and standard error via nested cross-validation.

Main Results:

• DL models achieved optimal mean absolute errors between 5.22 and 6.06 dB across tested frequencies.
• Standard errors ranged from 7.27 to 8.55 dB.
• All developed DL models demonstrated superior performance compared to traditional ML models.

Conclusions:

• Swept-tone SFOAE-based DL models can quantitatively predict hearing thresholds with satisfactory accuracy.
• DL techniques effectively capture the complex relationship between SFOAEs and hearing thresholds.
• These findings suggest potential for enhanced diagnostic value of SFOAEs in audiology.