Automatic analysis of cochlear response using electrocochleography signals during cochlear implant surgery

Abstract

Cochlear implants (CIs) provide an opportunity for the hearing impaired to perceive sound through electrical stimulation of the hearing (cochlear) nerve. However, there is a high risk of losing a patient's natural hearing during CI surgery, which has been shown to reduce speech perception in noisy environments as well as music appreciation. This is a major barrier to the adoption of CIs by the hearing impaired. Electrocochleography (ECochG) has been used to detect intra-operative trauma that may lead to loss of natural hearing. There is early evidence that ECochG can enable early intervention to save natural hearing of the patient. However, detection of trauma by observing changes in the ECochG response is typically carried out by a human expert. Here, we discuss a method of automating the analysis of cochlear responses during CI surgery. We establish, using historical patient data, that the proposed method is highly accurate (∼94% and ∼95% for sensitivity and specificity respectively) when compared to a human expert. The automation of real-time cochlear response analysis is expected to improve the scalability of ECochG and improve patient safety.

Fig 1

Top panels: left shows the stimulus of alternating polarity and right shows the raw responses recorded for both condensation and rarefaction stimuli. Bottom panels: DIF response calculated by subtracting rarefaction and condensation responses is on the left while the SUM calculated by the addition of the responses is on the right.

Fig 2. Changes in the cochlear microphonic for two patients with presumed electrophysiological evidence of atraumatic (left) and traumatic (right) CI insertions.

Fig 3. Results of the real-time peak detection algorithm.

All peaks of the CM signal are shown as red dots. The peaks detected by the algorithm as those an expert would pick in determining drops are shown in blue circles.

Fig 4. Results of the post-processing algorithm.

The left panel shows the classification results prior to post-processing. The right panel shows how the post-processing algorithm has removed some of the misclassifications. Drop points detected by the human expert are given as red dots while those detected by the automated algorithm are given in blue circles.

Fig 5. Components used in the calculation of performance metrics.

B-C and H-I are True ‘Drops’. L-M are False ‘No Drops’. D-E and J-K are False ‘Drops’. A and F-G are drops identified before the human expert and therefore are considered to be True ‘Drops’. All other instances are True ‘No Drops’.

Fig 6. Correlation of feature pairs as measured using Pearson’s correlation coefficient.

Yellow indicates perfect correlation (on the diagonal between the same features). Green, light blue, and dark blue show the pairs with moderate, low, and negligible correlations respectively.

Fig 7. Effect of drop misclassification cost on performance results.

Changes in sensitivity and specificity when the drop misclassification cost is increased is shown in blue and red respectively. The black dashed line indicates the cut-off level of 0.9 which was deemed acceptable for both metrics.

Fig 8. Feature importance in drop classification.

The feature numbers refer to those defined in Eq 1. The horizontal axis shows the features sorted in descending order of importance (or weight). The vertical axis shows the cumulative sum of the feature weights.