Background & Summary

Electrocochleography (ECochG) measures electrophysiological inner ear potentials in response to acoustic stimulation. These potentials reflect the state of the inner ear and provide important information about its residual function. ECochG is an umbrella term covering four different signal components, i.e., i) the cochlear microphonic (CM, outer hair cell response), ii) the auditory nerve neurophonic (ANN, early neural and inner hair cells response), iii) the compound action potential (CAP, early auditory nerve response), and iv) the summating potential (SP, mainly inner hair cell response)1,2,3,4,5.

In cochlear implant (CI) patients, using the implant electrode, we can measure ECochG signals directly within the cochlea. The measurements can be performed during and after implantation. During the implantation process, studies have shown that abrupt signal changes can be caused by traumatic forces6,7,8,9,10,11,12,13,14. Hence, real-time ECochG traces can complement the haptic perception of the surgeon6,8,9,10,11,15,16,17,18,19,20. ECochG can also be useful in the post-operative phase, where patients may lose residual cochlear function21,22. Most commonly, such losses occur during the first six to twelve months after implant surgery23,24,25 due to different intra-cochlear factors (e.g., immune response to the electrode, intracochlear inflammatory reactions, and intracochlear scar tissue formation)14,26,27. However, the underlying mechanisms remain poorly understood and require further research24. In summary, in CI recipients, during and after implant surgery, ECochG measurements map cochlear health and thus have great potential to improve our understanding of cochlear function in response to the implant electrode.

The interpretation of ECochG signals, however, is not trivial and requires training. The signal amplitude and signal-to-noise ratio (SNR) can vary greatly among individuals. Furthermore, the morphology and latency of ECochG traces are affected by the remaining neurosensory cells10,28,29,30.

Until recently, the evaluation of ECochG signals was based on visual analysis by experts. This approach has several disadvantages, e.g., a high level of experience is needed, and expert-dependent analysis can lead to a lack of reproducibility, limiting the application of these measurements. We previously introduced a machine learning-based, objective method to determine whether an ECochG signal is present or not31. Thereby, three experts labelled more than 4,000 ECochG signals to train and test the machine learning algorithm (consisting of preprocessing steps and a convolutional neural network, CNN).

The aim of this data descriptor is to provide the research community access to our comprehensive electrophysiological data set and algorithms (i.e., raw data with access down to single epoch level, pre-processing and SNR enhancing algorithms, visually labeled data by three independent human experts, and the trained deep learning network AlexNet)31. These data are complemented by the measured hearing thresholds, subjective loudness data, demographic data, impedance telemetry measurements, and radiographic parameters.

Potential applications of this data set include, but are not limited to (i) refinement and further use of the deep learning network31, (ii) improvement of pre-processing and SNR-enhancing algorithms and data analysis16,31,32,33, (iii) correlation of ECochG signal components and impedance measurements with hearing thresholds15,16,21,22,34, (iv) longitudinal evaluation and repeatability assessment of ECochG data21, and (v) correlation of multi-frequency and broad-band ECochGs with pure tone ECochGs and hearing thresholds35.

Methods

The data presented in this descriptor were collected in a study that was approved by our local institutional review board (The Cantonal Ethics Committee of Bern, BASEC ID 2019-01578). All participants gave written consent and consent to the use of properly anonymized data before participation.

Subject demographics

We recorded ECochG traces from 41 adult subjects (n = 46 ears) using a cochlear implant (MED-EL, Innsbruck, Austria). The subjects’ mean age was 58 years (SD = 17.4 yrs, range: 21 to 86 yrs). Pure tone audiograms were performed in a certified acoustic chamber with a clinical audiometer (Interacoustics, Middelfart, Denmark). Hearing thresholds were collected either immediately pre-operatively (cohort A) or, in the case of post-operative measurements (cohort B), on the day of ECochG measurement. We obtained pure tone air conduction hearing thresholds in dB hearing level (HL) at 125, 250, 500, 750, 1000, 1500, 2000, and 4000 Hz. For cohort A, we only included subjects with a hearing threshold at 500 Hz of 100 dB hearing level (HL) or better. For cohort B, we only considered subjects with stable acoustic hearing six months or longer after the implantation. The acoustic hearing was considered stable if the hearing thresholds varied less than 10 dB. In cohort B, subjects categorized the loudness of the acoustic stimulus according to Fig. 136.

Fig. 1
figure 1

Categories of subjective loudness. Subjects from cohort B classified each acoustic stimulus intensity to one of these categories.

Full size image

ECochG data

ECochG recordings were performed using MED-EL Maestro research software (versions 8.03 AS and 9.03 AS). The acoustic stimulus was generated by a Dataman 531 waveform generator (Dataman, Maiden Newton, UK) and converted to sound by an Etymotic ER-3C transducer (Etymotic, Grove Village, IL, USA). The acoustic stimulus was triggered via the MED-EL MAX interface. Further details are available in19.

We measured ECochG signals in response to pure tone, click, and SPL chirp stimulus (see Table 1 and Fig. 2). We recorded two polarities (condensation, CON, and rarefaction, RAR) and 100 repetitions (epochs) each. All ECochG recordings were measured in a stable electrode position; either in the operating room after completed electrode insertion (cohort A, 25 ears, the measurement setup can be found in19,37) or in a post-operative setting (cohort B, 21 ears) in a certified acoustic chamber. We thereby measured ECochG traces at electrodes 1 (most apical electrode), 4, 7, and 10 and in response to 3 different sound intensity levels (supra-threshold level, near-threshold level, sub-threshold level). The intensity levels were calculated using the individual hearing thresholds measured before the experiment. Our goal was to evoke responses with different SNRs. For cohort B, to obtain longitudinal data, we repeated ECochG recording three times: i) at least 6 months after insertion; ii) within 2 to 48 hours after the first measurement; and iii) 2 to 4 months after the first measurement.

Table 1 Table 1 shows the different stimulation modalities with the three stimuli used, the frequencies, the stimulus duration, and the measurement window.
Full size table
Fig. 2
figure 2

Electrical signals (left) and acoustic signals (right) generated by the waveform generator and transducer, respectively: A) 500 Hz pure tone, B) click, C) SPL chirp v1, and D) SPL chirp v2 stimulus. Note the different time axes (X-axis) scaling. The amplitude axes (Y-axis) were normalized. Acoustic signals were measured using a head and torso simulator (Type 5128-C-111, Brüel & Kjær, Virum, Denmark) and an audio analyzer (XL2, NTi Audio AG, Schaan, Lichtenstein). Electrical signals were measured using an oscilloscope (TDS 1002B, Tektronix, Beaverton OR, USA).

Full size image

Data preprocessing

To pre-process ECochG signals, we implemented the following steps (see31 for further details): i) if needed, removal of stitching artifacts; ii) application of a Gaussian-weighted averaging method adapted from33 to remove uncorrelated epochs; and iii) application of a second-order Butterworth band-pass filter in forward-backward filtering mode (cutoff frequencies at 10 Hz/5 kHz for visual analysis, and 100 Hz/5 kHz for the objective algorithms). The SNR was calculated using the ± averaging method38. The pre-processing steps above were performed using the Python script do_preprocessing.py, which is available at39.

Data analysis

For further analysis, we calculated the different ECochG signal components. We highlighted the CM signal by subtracting the CON and RAR responses40. Since the subtracted result can also contain other ECochG components, we will refer to the term “CM/DIF” signal in the following text32. We calculated the ANN signal by adding the ECochG response to CON and RAR stimulus3. For the following text, we will refer to it as “ANN/SUM” response.

For the visual analysis, the data were labeled by three independent experts with several years of experience in the field. Data were presented using Labelbox41 presenting a figure showing i) the CM/DIF trace, ii) the ANN/SUM trace, iii) the CON and RAR traces, and iv-vi) their corresponding Fast Fourier Transform (FFT) magnitude spectra. An example is shown in Fig. 3. During the labeling process, the focus lay on the identification of CM/DIF responses and their binary labeling (ECochG response visible/not visible). Thereby, the experts were forced to make a judgment; otherwise, it was not possible to proceed to the next signal trace. For the labeling of the ANN/SUM and CAP responses, however, in case of ambiguity, the answer could be skipped. The examiners did not discuss their evaluation to avoid bias in the assessment. Signals that were classified as visible CM/DIF responses by two examiners and as noise by the third examiner were presented a second time. Only, if all three experts rated a signal as visible (in the second round), it was marked as such. This was done to avoid volatility errors. Finally, we used the labeled responses to train the deep learning algorithm presented in31.

Fig. 3
figure 3

Visual analysis of ECochG traces was performed using six subplots. A) CM/DIF trace, C) ANN/SUM trace, E) CON and RAR traces, and B, D, F) their FFT traces. The gray vertical lines indicate the stimulus period. The dashed vertical lines indicate the expected frequency of the response.

Full size image

Impedance telemetry

Before each measurement session, we performed impedance telemetry measurements. We used the default settings of the recordings, recommended by the manufacturer. A charge-balanced, rectangular biphasic cathodic first pulse with a duration of 26.67 μs and an amplitude of 302.4 cu (one current unit, cu, is equivalent to approximately one μA) was used for stimulation resulting in a stimulation charge of 8.06 qu (one charge unit, qu, is equivalent to approximately one nC)42. The voltage potential was measured at the end of the anodic phase with respect to the ground electrode located at the implant housing43,44.

Anatomy

Anatomical features were extracted from the Computed tomography (CT) scans using Otoplan (ver. 1.02, CAScination, Bern, Switzerland)45. CT images with a slice thickness equal to or less than 0.3 mm were used. Markers to define the cochlea were set (A value, distance between the round window and the contralateral wall of the cochlea, B value, width of the cochlea perpendicular to the A value, H value, distance from the basal turn to the apical center)46,47.

Data Records

All data created during this research project are accessible from the Dryad repository39. The dataset is stored in the Bern ECochG SQL database, and consists of seven tables, as shown in Fig. 4. Each table can be accessed individually. All tables except the Analysis table use the common Subject id attribute, which can be employed to connect the tables.

Fig. 4
figure 4

The Bern ECochG database contains seven tables.

Full size image

Subject demographics

The subject’s demographic data is stored in the Demographics table. A list of all attributes is available in Table 2. The Subject id is stored as XX_Y, where XX is post-insertion (PI) or post-operative (PO) and Y is an incrementing number for each subject. The Python script demographics.py illustrates how to access the demographic data.

Table 2 Structure of the subject Demographics table.
Full size table

Hearing thresholds

The subjects’ hearing thresholds are stored in table Hearing thresholds. A list of all attributes can be found in Table 3. For cohort A, we provide immediate, pre-operative and 3–5 weeks post-operative hearing thresholds. For cohort B, we list the hearing threshold before the first post-operative ECochG recording (post-operative) and before the third post-operative recording (post-operative 2). In case of a missing hearing threshold, we left cells blank.

Table 3 Structure of the Hearing thresholds table.
Full size table

ECochG data

The table ECochG contains all ECochG raw data. A list of all attributes can be found in Table 4. The measurement date shows when the measurement was performed. Measurement session indicates to which session the measurement belongs (0: post-insertion, cohort A, 1–3: post-operative measurements, cohort B). Measurement number is an ascending number for each session. stimulus type indicates which acoustic stimulus was used for the recording. Stimulus duration indicates the duration of the acoustic stimulus in milliseconds (ms). Polarity indicates whether a CON or RAR stimulus was used. The acoustic amplitude of the stimulus is given in dB hearing level (dB HL) for pure tones or in dB peak equivalent sound pressure level dB p.e. SPL for click and SPL chirp stimulus29. The Recording window indicates the length of the recording in ms. The Measurement delay specifies the delay between the start of the acoustic stimulus and the start of the measurement window. In most cases, Measurement delay is set to 1 ms. Timeaxis and Signal are Numpy arrays stored as JSON strings48,49. The Timeaxis was stored as a 1 × N array, where N indicates the time samples. The Signal was stored as M × N, where M indicates the recorded epochs and N indicates the recording samples. Subjective loudness represents the loudness of the acoustic stimulus as perceived by the subjects (cohort B). Available responses are shown in Fig. 1.

Table 4 Structure of the ECochG table.
Full size table

Preprocessed

The Preprocessed table contains data generated after the pre-processing steps. The attributes are listed in Table 5. The signal is indicated by s.

Analysis

The Analysis table contains the visual and objective analysis of the signals. The analyzed signals consist of a pair of CON and RAR recordings. The recordings can be traced using the Id, which is represented as XX_Y.SESSION_NR.NR_CON.NR_RAR. Where XX is PO or PI, Y represents the subject’s incrementing identification number, SESSION_NR is the session number, and NR_CON and NR_RAR represent the measurement numbers (e.g., PO_1.1.010.011 consists of recordings #10 and #11 of post-operative subject 1 and session 1, respectively).

Table 5 Structure of the Preprocessed table.
Full size table

Analysis was performed for CM/DIF (DIF), ANN/SUM (SUM), and CAP components. ECochG components were labeled by the examiners (l1 - l3) and the deep learning (DL) algorithm.

Objective analysis of the CM/DIF signals is only available for pure tone stimulus. Unlabeled components were left blank. Table 6 shows an overview of all attributes available in the Analysis table.

Table 6 Structure of the Analysis table.
Full size table

Anatomy

The Anatomy table contains the anatomical features. A list of all attributes can be found in Table 7. Type indicates whether the anatomical features were extracted from pre-operative or post-operative CT images. The shape of the cochlea is indicated by the A, B, and C values, and the cochlear duct length (CDL)46. General statistics about the anatomical features are shown in Table 8

Table 7 Structure of the Anatomy table.
Full size table
Table 8 General statistics about the anatomical features extracted from CT scans.
Full size table

Impedance telemetry

The table Telemetry contains the recorded values during clinical routine telemetry measurements. A list of all attributes can be found in Table 9. The Clinical impedances represent the impedances from the electrodes (1 to 12) to the ground electrode. General statistics about clinical impedances are shown in Table 10

Table 9 Structure of the Telemetry table.
Full size table
Table 10 General statistical data on impedance measurements with electrode one as the most apical.
Full size table

Technical validation

The ECochG system was calibrated by the manufacturer. No changes were made to the recorded raw data. To increase the reliability of the measurements in cohort A, we used sterile eartips for recording and applied the guidelines presented in37. In cohort B, we compared hearing thresholds measured with the ECochG hardware with the audiogram before each measurement session. In this way, we were able to verify that the eartips were placed correctly. For this purpose, we used the customized software AcousticStimulatorGUI, available from39. This software interacts directly with the Dataman waveform generator and allows the use of customized acoustic stimulus. The software with the corresponding hardware was calibrated on a head and torso simulator (Brüel & Kjær, type 5128, Nærum, Denmark). The AcousticStimulatorGUI was calibrated with our hardware. Using this software together with other hardware requires a new calibration. The calibration parameters can be adjusted in the GetFrequencyOffset method of the Dataman class.

Usage Notes

The database has been split into seven data parts and the empty Bern_ECochG database to facilitate downloading. Each part is saved as a .sql file and can be imported into the Bern_ECochG database individually. We recommend downloading all parts and assembling them using sqlitebrowser available at https://sqlitebrowser.org/. The Python scripts provided will only work when the database is fully assembled. The Python scripts show how to access the database. Along the Python scripts, a .yml file is provided to install all dependencies to run the scripts.