Deep learning models for predicting hearing thresholds based on joint stimulus-frequency otoacoustic emissions and distortion-product otoacoustic emissions

Abstract

According to the dual-source generation hypothesis, stimulus-frequency otoacoustic emissions (SFOAEs) and distortion-product OAEs (DPOAEs) arise from different cochlear mechanisms, and both are capable of characterizing hearing loss. However, their joint application for hearing threshold prediction remains unexplored. This study developed an efficient deep learning (DL) model integrating SFOAEs and DPOAEs to quantitatively predict hearing thresholds. Training data for the model were collected from 94 ears with normal hearing and 401 ears with sensorineural hearing loss. Frequency-specific DL models were constructed across five octave frequencies (0.5–8 kHz), with inputs including amplitude spectra and corresponding signal-to-noise ratio spectra of both SFOAEs and DPOAEs. Self-extractors of the model were constructed using convolutional neural network (CNN) and recurrent neural network (RNN), respectively. Cross-validation demonstrated that the dual-OAE model achieved mean absolute errors (MAEs) of 5.17, 3.83, 3.96, 4.71, and 4.90 dB at 0.5–8 kHz, significantly outperforming single-OAE DL models (except DPOAE-based models at 0.5 and 2 kHz) and baseline machine learning models. By reducing the number of OAE stimulus levels, the efficiency-optimized model reduced testing time per individual to approximately 15 min while preserving accuracy. The proposed dual-source OAE (SFOAE and DPOAE)-integrated DL model achieves state-of-the-art accuracy in hearing threshold prediction, with its optimized efficiency establishing a foundation for further developing a practical clinical tool for objective hearing loss diagnosis.