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

Abstract

Objectives: This study aims to develop deep learning (DL) models for the quantitative prediction of hearing thresholds based on stimulus-frequency otoacoustic emissions (SFOAEs) evoked by swept tones. Design: A total of 174 ears with normal hearing and 388 ears with sensorineural hearing loss were studied. SFOAEs in the 0.3 to 4.3 kHz frequency range were recorded using linearly swept tones at a rate of 2 Hz/msec, with stimulus level changing from 40 to 60 dB SPL in 10 dB steps. Four DL models were used to predict hearing thresholds at octave frequencies from 0.5 to 4 kHz. The models—a conventional convolutional neural network (CNN), a hybrid CNN-k-nearest neighbor (KNN), a hybrid CNN-support vector machine (SVM), and a hybrid CNN-random forest (RF)—were individually built for each frequency. The input to the DL models was the measured raw SFOAE amplitude spectra and their corresponding signal to noise ratio spectra. All DL models shared a CNN-based feature self-extractor. They differed in that the conventional CNN utilized a fully connected layer to make the final regression decision, whereas the hybrid CNN-KNN, CNN-SVM, and CNN-RF models were designed by replacing the last fully connected layer of CNN model with a traditional machine learning (ML) regressor, that is, KNN, SVM, and RF, respectively. The model performance was evaluated using mean absolute error and SE averaged over 20 repetitions of 5 × 5 fold nested cross-validation. The performance of the proposed DL models was compared with two types of traditional ML models. Results: The proposed SFOAE-based DL models resulted in an optimal mean absolute error of 5.98, 5.22, 5.51, and 6.06 dB at 0.5, 1, 2, and 4 kHz, respectively, superior to that obtained by the traditional ML models. The produced SEs were 8.55, 7.27, 7.58, and 7.95 dB at 0.5, 1, 2, and 4 kHz, respectively. All the DL models outperformed any of the traditional ML models. Conclusions: The proposed swept-tone SFOAE-based DL models were capable of quantitatively predicting hearing thresholds with satisfactory performance. With DL techniques, the underlying relationship between SFOAEs and hearing thresholds at disparate frequencies was explored and captured, potentially improving the diagnostic value of SFOAEs.