Classification of Acoustic Hearing Preservation After Cochlear Implantation Using Electrocochleography

Abstract

The objective to preserve residual hearing during cochlear implantation has recently led to the use of intracochlear electrocochleography (ECochG) as an intraoperative monitoring tool. Currently, a decrease in the amplitude of the difference between responses to alternating-polarity stimuli (DIF response), predominantly reflecting the hair cell response, is used for providing feedback. Including other ECochG response components, such as phase changes and harmonic distortions, could improve the accuracy of surgical feedback. The objectives of the present study were (1) to compare simultaneously recorded stepwise intracochlear and extracochlear ECochG responses to 500 Hz tone bursts, (2) to explore patterns in features extracted from the intracochlear ECochG recordings relating to hearing preservation or hearing loss, and (3) to design support vector machine (SVM) and random forest (RF) classifiers of acoustic hearing preservation that treat each subject as a sample and use all intracochlear ECochG recordings made during electrode array insertion for classification. Forty subjects undergoing cochlear implant (CI) surgery at the Oslo University Hospital, St. Thomas' Hearing Implant Centre, or the University Hospital of Zurich were prospectively enrolled. In this cohort, DIF response amplitude decreases did not relate to postoperative acoustic hearing preservation. Exploratory analysis of the feature set extracted from the ECochG responses and preoperative audiogram showed that the features were not discriminative between outcome classes. The SVM and RF classifiers that were trained on these features could not distinguish cases with hearing loss and hearing preservation. These findings suggest that hearing loss following CI surgery is not always reflected in intraoperative ECochG recordings.

Keywords: classifier; cochlear implant; electrocochleography; random forest; residual hearing; support vector machine.

Figure 1.

Development cycle of the ECochG-based SVM and RF for classifying cochlear trauma. Scaling is only applied as part of step 3 for SVM development. ECochG=electrocochleography; RF=random forest; SVM=support vector machine.

Figure 2.

Overview of the process used to train and evaluate the SVM and RF models. Initially, the dataset is divided into a training and test set. The training set is then split into five folds. For each combination of hyperparameters, five model fits are made, whereby four of the folds serve as a training set and one fold serves as a validation set. The resulting five f1 scores are averaged and the model with the highest mean f1 (training score) is chosen. The performance of the final model is evaluated using the held-out test dataset. RF=random forest; SVM=support vector machine.

Figure 3.

The preoperative and postoperative audiograms (mean ± SD) for subjects with hearing preservation and hearing loss. The audiograms of the child in whom postoperative audiometry was only conducted at 0.5 kHz, and who had a partial hearing loss on both sides, were omitted from the figure.

Figure 4.

Amplitude and phase component at 500 Hz of the extracochlear and intracochlear DIF responses for cases with sudden, near 180-degree phase shifts (shown by the arrows). The phases determined are not corrected for the cycle. Measurements within the noise floor are represented by open symbols, while measurements above the noise floor are represented by filled symbols. Hearing was preserved in S03 and S06 (hearing preservation (HP), green), while it was partially lost in S07 (partial hearing loss (PHL), orange) and completely lost in S04 and S09 (complete hearing loss (CHL), red).

Figure 5.

Amplitude and phase component at 500 Hz of the extracochlear and intracochlear DIF responses for cases with slight, gradual phase changes. The phases determined are not corrected for the cycle. Measurements within the noise floor are represented by open symbols, while measurements above the noise floor are represented by filled symbols. Hearing was partially preserved in S01, S02, and S08 (partial hearing loss (PHL), orange). Measurements of S05 and S10 were excluded because the initial recordings were within the noise floor.

Figure 6.

DIF and SUM response amplitudes (mean ± SD) of the continuous ECochG recordings for subjects with and without hearing preservation, plotted against insertion depth. Only recordings that were included for classification are shown. ECochG=electrocochleography.

Figure 7.

Features 5, 6, 9, 12, 13, and 14 versus the maximum DIF response amplitude decrease (feature 4), for both the hearing preservation and the hearing loss class.

Figure 8.

Maximum SVM (left) and RF (right) cross-validation f1-score for the training iterations using Bayes optimization, random search, and grid search. RF=random forest; SVM=support vector machine.