Preoperative prediction of angular insertion depth of lateral wall cochlear implant electrode arrays

Abstract.

Purpose: Cochlear implants (CIs) use an array of electrodes surgically threaded into the cochlea to restore hearing sensation. Techniques for predicting the insertion depth of the array into the cochlea could guide surgeons toward more optimal placement of the array to reduce trauma and preserve the residual hearing. In addition to the electrode array geometry, the base insertion depth (BID) and the cochlear size could impact the overall array insertion depth.

Approach: We investigated using these measurements to develop a linear regression model that can make preoperative or intraoperative predictions of the insertion depth of lateral wall CI electrodes. Computed tomography (CT) images of 86 CI recipients were analyzed. Using previously developed automated algorithms, the relative electrode position inside the cochlea was measured from the CT images.

Results: A linear regression model is proposed for insertion depth prediction based on cochlea size, array geometry, and BID. The model is able to accurately predict angular insertion depths with a standard deviation of 41 deg and absolute deviation error of 32 deg.

Conclusions: Surgeons may use this model for patient-customized selection of electrode array and/or to plan a BID for a given array that minimizes the likelihood of causing trauma to regions of the cochlea where residual hearing exists.

1. Introduction

Cochlear implants (CIs) are considered to be an effective treatment for severe-to-profound hearing loss.¹ In CI surgery, an electrode array is implanted in the cochlea to directly stimulate the auditory nerve (see Fig. 1). In natural hearing, a neural pathway gets activated when the incoming sound has the characteristic frequency of that pathway. These pathways are arranged tonotopically by decreasing frequency along the length of the cochlear duct.² The implanted electrode arrays are used to stimulate the nerve pathways and induce hearing sensation.³ The CI processor receives and processes the incoming sound into signals sent to the electrodes. CI programming parameters, defined by an audiologist, determine which electrodes are triggered based on the frequencies present in the received sound.³ CIs leverage the natural tonotopy of the nerve pathways by assigning lower frequency sounds to stimulate deeper electrodes on the array and vice versa.

Fig. 1.

Insertion depth of the lateral wall-positioned straight array is measured in the angle around the midmodiolar axis to the most distal contact (540 deg).

Insertion depth of the electrode array into the cochlea has been an often studied factor when attempting to explain variability in speech perception outcomes with CIs. For example, some studies find depth of insertion to be positively correlated with speech perception^4–7 while other studies report data supporting no relationship^8,9 and others supporting a negative correlation.^10,11 Heutink et al.¹² mentioned that the impact of the insertion depth on speech perception does not have firm support according to the current body of evidence. Our recent work suggests that this relationship may depend on the type of CI electrode array that is used.¹³ Thus, conflicting evidence is available in the general cases of CI recipients with no residual hearing. However, a growing population of CI recipients are individuals who have residual hearing where lower frequency (deeper) hair cells and neural pathways are still functional. For that population, array placement in functional regions could negatively impact residual hearing. The lateral and modiolar walls of the cochlea refer to the outer wall of the cochlear spiral and the inner wall closest to the midmodiolar axis, respectively (see Fig. 1). With straight electrode arrays (a.k.a. “lateral wall” arrays), surgeons push the array into the cochlea through a small membrane called the “round window” (RW) at the base of the cochlea while the tip of the array is guided deep into the cochlea by sliding along the lateral wall. Therefore, the lateral wall electrode arrays are likely to be placed adjacent to the lateral wall in contrast with the perimodiolar and precurved electrode arrays, which are intended to hug the modiolus. For residual hearing subjects implanted with lateral wall arrays, the surgeon’s goal is to insert the array as deep as possible so that greater coverage of the auditory nerve is achieved, but not so deep as to cause trauma in the deeper regions where healthy hair cells exist. Lacking visualization into the cochlea, the surgeon does not know how deep the electrode array is being inserted on a patient-specific basis, making it difficult to avoid potential damage to residual hearing regions. Thus, tools that can help surgeons avoid overinserting and damaging healthy regions are beneficial. Here, we propose an approach for preoperative prediction of angular insertion depth based on the implant length and diameter, cochlear scale, and the base insertion depth (BID). Cochlear scale is a metric that aims to capture the size of the cochlea. The BID is a metric that represents the linear depth of the base of the electrode array that is inserted into the cochlea, which is controllable and observable by the surgeon. The BID together with the array length account for the total length that the array is inserted into the cochlea, while the array diameter and cochlear scale account for how long or short of a path the array takes within the cochlea to reach a given angular insertion depth. Using the angular insertion depth prediction that we propose, it is possible for the surgeon to preoperatively plan a customized BID for a given array to achieve a desired overall angular insertion depth of the array.

Cochlear anatomy, including the overall size and the number of cochlear spiral turns, varies significantly in humans.^14–16 From the computed tomography (CT) images of the temporal bones of patients having congenital hearing defects, it is reported that about 25% of these patients have bony inner ear deformities.¹⁷ Even when limited to normal cochleae, the length of the organ of Corti is variable among individuals.^14,15,18,19 In prior studies, the length of the lateral wall covered by an electrode array is defined as the cochlear coverage length (CCL), and the ratio of the CCL and the cochlear duct length (CDL) is considered to be the cochlear coverage (CC).²⁰ Timm et al.²⁰ showed that the variation of the CDL among patients impacts the electrode specific CC. The angular insertion depth has a strong correlation with CC ($R=0.98$).²¹ Several studies have been carried out to establish the relationship between the angular insertion depth and the cochlear shape. In one study, the angular insertion depth and its relationship with the cochlear size and the linear insertion depth were investigated in the case of a specific implant (Cochlear CI422).²² For different linear insertion depths, the study showed that there is a significant correlation between the angular insertion depth and the cochlear diameter. In the case of a 25-mm linear insertion depth, the study showed the most correlation between cochlear diameter and angular depth ($R^2=0.97$, $p<0.0001$) relative to other linear depths. In addition, for a 23-mm linear insertion depth, the correlation was also significant ($R^2=0.80$, $p<0.0001$)²². A linear insertion depth prediction model based on a targeted angular insertion depth was presented by Anschuetz et al.²³ Rathgeb et al.²⁴ evaluated the clinical applicability of the angular insertion depth prediction method by analyzing 50 cases of CI recipients in which all of the patients received the same (MED-EL Flex28) electrode array. However, the prediction model used in both studies did not consider the impact of the variation of the electrode array length and diameter of the electrode array on the angular insertion depth. Schurzig et al.²⁵ presented a geometric model to predict the angular insertion depth in the case of MED-EL Flex electrode arrays.

Escudé et al.¹⁷ studied the influence of the cochlear diameter upon the angular insertion depth, and Alexiades et al.²⁶ estimated CDL to select the array with a length that is most suitable for the cochlea. The largest distance from the center of the RW to the lateral wall through the modiolar axis, termed by Escudé et al.¹⁷ as “$A$,” was proposed as a metric for cochlear diameter that is easy to manually measure in CT images. An automatic method of measuring $A$ using an active shape model-based technique was presented by Rivas et al.²⁷ Escudé et al.¹⁷ showed that for a fixed linear insertion depth (17 mm) the angular insertion depth demonstrates a significant negative correlation ($R^2=0.51$, $p<0.05$) with the $A$ length. Adunka et al.²⁸ studied eight CI cases in which the MED-EL (Innsbruck, Austria) standard electrode array was used for assessing the feasibility and accuracy of estimating the insertion depth. However, these studies did not directly relate BID to angular depth. In this study, we investigated the angular insertion depth for various commonly used lateral wall electrode arrays (MED-EL Standard, Flex28, Flex24, Cochlear CI522, and Advanced Bionics HiFocus SlimJ). The main contributions of the current study are twofold. First, we present data that further demonstrates how the cochlear scale, implant geometry, and BID relate to the angular insertion depth of CI electrodes. Second, we present a data-driven model to predict the angular insertion depth for preoperative prediction and surgical planning. As opposed to the prediction models discussed above, which estimate depth using idealized geometric models, the approach we present herein builds a linear regression model from a retrospective dataset to facilitate a data-driven prediction of angular insertion depth for a given array and cochlea. We hypothesize that such a data-driven prediction may be more accurate than predictions from a geometric model. Such a prediction model can be used by the surgeon for selecting an array and planning a BID customized for individual patients, so the largest region of the auditory nerve can be stimulated while still avoiding overinsertion of the array into sensitive residual hearing regions.

2. Methodology

In this paper, we used a dataset of preoperative and postoperative CT images of 86 CI users who had lateral wall positioned straight electrode array CIs. Informed consent was obtained according to protocols approved by Vanderbilt IRB #090155. The dataset included individuals implanted with the current most commonly used straight electrode array devices, including the MED-EL Standard, Flex28, and Flex24; the Advanced Bionics SlimJ; and the Cochlear CI522 devices. The number of each device included in our dataset and the estimates of their length and width that we have used in this study are shown in Table 1. Length and width are available from publicly available datasheets provided by the manufacturers. In the case of tapered arrays, we have estimated the average diameter of the array based on these datasheets.

Table 1.

Different electrode array types and their corresponding lengths and diameters in our dataset.

Electrode array type	Number of cases	Array length (mm)	Array diameter (mm)
MED-EL (Standard)	5	31.5	0.75
MED-EL (Flex28)	31	28	0.60
MED-EL (Flex24)	1	24	0.50
Cochlear (CI522)	16	23	0.40
Advanced Bionics (HiFocus SlimJ)	33	23	0.55

A nonrigid shape model of intracochlear structures was used to segment the intracochlear structures in the preoperative CT image²⁹ or a synthetic preoperative CT generated using the postoperative CT using deep learning techniques to erase the artifact created by the presence of the implant.³⁰ The shape model was defined using high-resolution $\mu \mathrm{CT}$ images initially²⁹ of 6 specimens, but later expanded to 16 specimens. Automated algorithms were used to nonrigidly register the model to new patients’ CTs.^29,30 This automated localization process using the $N=16$ model has mean surface localization errors of $\sim 0.1\mathrm{mm}$ (unpublished observations). The localization process relies on active shape model methods to produce a mesh representing the scala tympani (ST). Because the ST is segmented using an active shape model approach, a one-to-one point correspondence is maintained across each case, so measurements of distances between landmarks, such as the $A$ length measurement of cochlear diameter or other measures of cochlear size, can be easily performed in an automated fashion. Our proposed metric of cochlear scale based on the ST segmentation will be described below.

In postoperative CT images, the electrodes were localized using automated methods that have been shown in validation studies to result in average localization errors of the centroid of each contact of $\sim 0.12\mathrm{mm}$.³¹ The average localization error was obtained by comparing automatic localizations in CT images of cochlea specimens with a ground truth found from high-resolution $\mu \mathrm{CT}$ scans of the same specimens. The electrode localization process starts with finding a set of candidate points for electrodes from the postoperative CT images. A curve is formed using the most probable candidate points for representing the electrode array.^31,32 To get the relative positions of the electrodes inside the cochlea, a rigid coregistration between the postoperative and preoperative CT images was performed.

Using the cochlea segmentation and the positions of the electrodes, we measured the array angular insertion depth, BID, and cochlear scale. Angular depth was determined as the insertion depth angle relative to the RW membrane of the most apical contact on the array. BID was measured as the distance from the most basal electrode to the RW membrane where the array is inserted into the cochlea. This is a signed distance, where positive distances indicate the most basal electrode is within the cochlea and negative distances indicate it is outside the cochlea. This metric was chosen relative to the most basal electrode, rather than the surgical depth marker, because the depth marker is not identifiable in a CT scan for most electrode types. However, the most basal electrode is identifiable and is close enough (3 to 5 mm) to the depth marker that the variability seen in the BID measure we use should be the same as the variability in the position of the marker relative to the RW for the vast majority of cases where there are no kinks in the array and it is relatively straight between the marker and basal electrode. The cochlear scale metric we propose is computed as

$$\text{Cochlear scale}=\frac{1}{N}\sum _{i=1}^N|\overrightarrow{v_i}-\overrightarrow{c}|,$$ (1)

i.e., it is the average distance of the set of all $N$ vertices of the ST mesh, ${\{{\overrightarrow{v}}_i\}}_{i=1}^N$, to the centroid of that set of vertices, $\overrightarrow{c}=\frac{1}{N}\sum _{i=1}^N{\overrightarrow{v}}_i$. In contrast with other metrics that are commonly used as surrogates for cochlear size such as the $A$ and $B$ lengths,^17,26 which are based on one or two distance measurements between different points on the cochlea, our approach estimates cochlea size using a comprehensive set of measurements sampled over the entire structure, which we hypothesize might better correlate with angular depth of the electrode. Figure 1 shows an example case where BID and angular insertion depth can be seen for an electrode array. For this particular case, the BID is 2 mm and the angular insertion depth is 540 deg. Measurements for each case in our dataset are provided in Table 6 of the Appendix.

Table 6.

Overall dataset including base insertion depth, cochlear scale, $A$ length, $B$ length, array type, and actual and predicted angular insertion depths.

BID (mm)	Cochlear scale (mm)	$A$ length (mm)	$B$ length (mm)	Array type	Actual angular insertion depth (deg)	Predicted angular insertion depth (deg)
2.25	2.83	9.13	6.66	Flex28	581	524
2.45	2.70	8.88	6.49	Flex28	637	560
3.25	2.83	9.18	6.88	Flex28	489	551
0.75	2.76	8.97	6.79	Standard	637	610
3.65	2.94	9.55	6.96	Flex28	513	533
3.05	2.97	9.61	7.13	Flex28	449	512
3.35	2.95	9.47	7.11	Flex28	465	523
2.35	2.80	9.04	6.57	Flex28	545	536
2.10	2.73	8.89	6.51	Flex28	584	547
2.15	2.90	9.47	6.91	Flex28	454	508
2.85	2.71	8.86	6.43	Flex28	637	568
1.80	2.78	8.90	6.71	Standard	661	629
3.20	3.08	9.98	7.28	Standard	584	591
1.70	2.58	8.45	6.20	Flex28	600	574
3.00	2.82	9.08	6.79	Flex28	537	546
3.65	2.85	9.25	6.76	Flex24	461	456
$-0.68$	2.67	8.66	6.32	Flex28	500	495
3.20	2.53	8.16	6.11	Flex28	717	614
1.85	2.76	8.99	6.62	Flex28	516	536
3.25	2.75	8.86	6.56	Flex28	510	573
1.85	2.76	8.93	6.66	Flex28	483	535
0.45	2.91	9.48	7.09	Standard	584	566
1.40	2.81	9.08	6.67	Flex28	584	510
2.00	2.73	8.85	6.57	Flex28	523	547
0.70	2.58	8.48	6.08	Flex28	513	554
1.90	2.99	9.74	7.24	Flex28	502	475
$-2.06$	2.77	8.89	6.58	Flex28	416	442
$-3.44$	2.62	8.61	6.23	Flex28	394	450
$-1.33$	2.74	8.95	6.53	Flex28	478	461
0.10	2.77	9.03	6.67	Flex28	474	491
3.35	3.06	9.96	7.24	Flex28	527	492
$-2.63$	2.80	9.28	6.72	Standard	447	532
1.80	2.77	8.97	6.48	Flex28	554	529
$-0.85$	2.85	9.20	6.84	Flex28	426	449
1.90	2.98	9.47	7.25	Flex28	458	480
1.75	2.73	9.01	6.52	Flex28	603	537
4.40	2.81	9.09	6.75	Flex28	581	583
2.50	2.76	8.85	6.69	HiFocus SlimJ	472	448
$-0.98$	2.71	8.90	6.45	HiFocus SlimJ	366	379
1.75	2.81	9.19	6.77	HiFocus SlimJ	398	419
$-0.20$	2.97	9.65	7.10	HiFocus SlimJ	357	330
1.55	2.66	8.51	6.29	HiFocus SlimJ	480	448
3.17	2.79	9.00	6.82	HiFocus SlimJ	422	458
3.60	2.74	8.91	6.47	HiFocus SlimJ	419	483
0.70	2.79	8.96	6.76	HiFocus SlimJ	376	398
2.35	3.01	9.92	6.99	HiFocus SlimJ	374	383
$-0.64$	2.78	9.00	6.71	HiFocus SlimJ	336	370
4.69	2.81	9.15	6.71	HiFocus SlimJ	467	490
2.60	2.80	9.07	6.61	HiFocus SlimJ	432	442
2.85	2.94	9.57	7.02	HiFocus SlimJ	443	409
1.45	2.69	8.43	6.36	HiFocus SlimJ	450	441
$-0.28$	2.71	8.74	6.57	HiFocus SlimJ	391	396
$-0.39$	2.98	9.63	7.21	HiFocus SlimJ	311	325
$-1.95$	2.95	9.56	7.17	HiFocus SlimJ	398	280
1.10	2.64	8.52	6.36	HiFocus SlimJ	402	449
2.05	2.67	8.69	6.41	HiFocus SlimJ	538	456
0.55	2.68	8.66	6.47	HiFocus SlimJ	406	422
3.54	2.74	8.98	6.52	HiFocus SlimJ	487	477
2.00	2.88	9.05	6.88	HiFocus SlimJ	416	407
2.35	3.07	9.94	7.25	HiFocus SlimJ	365	366
2.70	2.85	9.29	6.78	HiFocus SlimJ	394	431
$-0.12$	2.92	9.56	6.96	HiFocus SlimJ	340	345
2.05	2.72	8.91	6.47	HiFocus SlimJ	431	449
$-0.96$	2.68	8.74	6.37	HiFocus SlimJ	374	387
4.57	3.09	9.94	7.37	HiFocus SlimJ	394	417
2.75	2.85	9.24	6.67	HiFocus SlimJ	417	433
2.90	3.03	9.97	7.13	HiFocus SlimJ	455	386
1.20	2.84	9.17	6.82	HiFocus SlimJ	388	398
2.25	2.90	9.44	6.96	HiFocus SlimJ	475	404
1.40	2.79	9.03	6.74	HiFocus SlimJ	371	415
$-0.32$	2.67	8.67	6.38	CI522	342	362
2.25	2.55	8.23	6.11	CI522	473	452
3.00	2.63	8.60	6.30	CI522	413	453
2.30	2.86	9.41	6.84	CI522	371	375
4.10	3.04	9.92	7.30	CI522	404	369
1.25	2.75	9.02	6.47	CI522	385	378
1.10	2.94	9.53	7.15	CI522	328	327
4.70	2.77	9.03	6.60	CI522	422	459
3.45	2.82	9.27	6.76	CI522	389	415
1.10	2.77	8.90	6.77	CI522	363	369
0.10	2.76	8.94	6.58	CI522	352	349
4.75	2.88	9.33	6.94	CI522	395	432
2.65	2.95	9.61	7.04	CI522	357	362
1.75	2.84	9.32	6.73	CI522	409	365
$-0.57$	2.83	9.16	6.71	CI522	359	310
3.60	2.82	9.30	6.72	CI522	422	416

For the prediction of angular insertion depth, a linear regression model was designed in this paper. The model was fitted and tested using a leave-one-out approach. In the leave-one-out approach, the model is trained to keep one data point out for testing and the process is repeated until every datapoint is used for testing. Here, we used 85 datapoints for fitting the regression model, and the remaining datapoint was used for testing the prediction performance of the model. This entire process was run 86 times, and each time a new datapoint was considered a testing point. The built-in MATLAB function fitlm was used to fit the linear regression model to the training data. We also trained individual models for each electrode array type listed in Table 1, with the hypothesis that array specific models may be more accurate than a general one.

3. Results and Discussion

The angular insertion depth among the 86 CI recipients ranged from 311 deg to 717 deg with the mean at 458 deg and standard deviation of 88 deg. In general, the standard and the Flex28 were associated with higher angular insertion depths compared with the other array models. The following sections detail the relationships observed between angular insertion depth and the other measured features.

3.1. Base Insertion Depth versus Angular Insertion Depth

The measurement of BID ranged from $-3.45$ to 4.75 mm with a mean of 1.73 mm and standard deviation of 1.73 mm. The histogram in Fig. 2 represents the distribution of the BID among the CI recipients. It is important to note that variability in BID is not equivalent to variability in how frequently surgeons follow or deviate from manufacturer recommendations. This is because the distance between the most basal electrode and the manufacturer marker varies across devices between 3 to 5 mm. Therefore, we would expect BID to range from 3 to 5 mm if the manufacturer recommended depth was used. It is noteworthy from Fig. 2 that the majority of cases have BID $<3\mathrm{mm}$, which indicates that in most cases array depth is more shallow than the manufacturer recommended depth.

Fig. 2.

Distribution of the base insertion depth.

The BID and the angular insertion depth did not show substantial direct correlation ($R^2=0.075$, $p=0.008$) because the cochlear scale and the electrode array length are not taken into consideration yet. However, after controlling for the other factors, the angular insertion depth showed strong correlation with the BID ($R^2=0.517$, $p=6.41e-15$). Figure 3 shows the variation in angular insertion depth with respect to the BID before and after controlling for the cochlear scale and the electrode array length and diameter. In cases of controlling for the other factors, first, BID is held out and the residual angular insertion depth is computed as the variance in the data unexplained by a linear regression model constructed with the remaining features. Similarly, the residual BID is computed as the variance in BID unexplained by a linear regression model created with the remaining features. Finally, the partial correlation between angular insertion depth and BID is the correlation between the residual angular insertion depth and residual BID. This is referred to as the correlation between two variables after controlling for the effect of other variables, and the same process is used below when referring to correlation after controlling for other variables.

Fig. 3.

Variation in angular insertion depth with respect to base insertion depth: (a) before controlling and (b) after controlling for the other factors.

3.2. Cochlear Scale versus Angular Insertion Depth

The cochlear scale, as calculated by Eq. (1), ranged from 2.53 to 3.09 mm with the mean at 2.81 mm and standard deviation of 0.13 mm. The histogram in Fig. 4 shows the distribution of the cochlear scale among the 86 CI recipients.

Fig. 4.

Distribution of the cochlear scale.

The effect of the cochlear scale on the angular insertion depth was observed with and without controlling the other factors. Figure 5 shows the variation in angular insertion depth with respect to the cochlear scale before and after controlling for the BID and electrode array geometry.

Fig. 5.

Variation in angular insertion depth with respect to cochlear scale: (a) before controlling and (b) after controlling for the other factors.

Without taking the BID and the array geometry into consideration, the cochlear scale and the angular insertion depth had weak correlation ($R^2=0.07$, $p=0.02$). After controlling for those factors, the angular insertion depth showed strong correlation with the cochlear scale ($R^2=0.39$, $p=1.29e-10$).

3.3. Electrode Array Length versus Angular Insertion Depth

The electrode array length and the angular insertion depth showed strong direct correlation ($R^2=0.55$, $p<0.001$). The partial correlation after controlling for the other factors was reduced but remained strong ($R^2=0.41$, $p=4.28e-11$). Figure 6 shows the variation in angular insertion depth with respect to the electrode array length before and after controlling for the BID, electrode array diameter, and the cochlear scale. The array length has strong correlation with the array diameter (see Table 2). Therefore, whenever we control for the array diameter, the residual correlation with length decreases.

Fig. 6.

Variation in angular insertion depth with respect to electrode array length: (a) before controlling and (b) after controlling for the other factors.

Table 2.

Correlation among angular insertion depth, BID, cochlear scale, electrode array length, and electrode array diameter.

	Angular insertion depth	BID	Cochlear scale	Electrode array length	Electrode array diameter
Angular insertion depth	1	0.27 [0.7, 0.46]	$-\mathrm{0.26}[-\mathrm{0.44},-\mathrm{0.05}]$	0.74 [0.63, 0.83]	0.62 [0.47, 0.73]
BID	0.27 [0.7, 0.46]	1	0.25 [0.04, 0.44]	$-0.10[-0.31,0.11]$	$-0.18[-0.37,0.04]$
Cochlear scale	$-0.26[-0.44,-0.05]$	0.25 [0.04, 0.44]	1	$-0.04[-0.25,0.17]$	0.03 $[-0.180.24]$
Electrode array length	0.74 [0.63, 0.83]	$-0.10[-0.31,0.11]$	$-0.04[-0.25,0.17]$	1	0.77 [0.66, 0.84]
Electrode array diameter	0.62 [0.47, 0.73]	$-0.18[-0.37,0.04]$	0.03 $[-0.180.24]$	0.77 [0.66, 0.84]	1

3.4. Electrode Array Diameter versus Angular Insertion Depth

The array diameter also plays a role in influencing the angular depth as electrodes with smaller diameters tend to have less offset from the lateral wall and take a longer path into the cochlea. Figure 7 shows the relationship of the angular insertion depth with the array diameter. A weak relationship was observed between the angular insertion depth and the array diameter ($R^2=0.15$, $p=3.01e-4$) after controlling for the other features because, as discussed above, array length and diameter are highly correlated.

Fig. 7.

Variation in angular insertion depth with respect to electrode array diameter: (a) before controlling and (b) after controlling for the other factors.

3.5. Summary of Relationships among Factors

Pearson’s correlation coefficients among these factors are provided in Table 2 followed by their 95% confidence intervals in square brackets.

It is evident that the angular insertion depth has statistically significant correlation with the BID, cochlear scale, array length, and array diameter, although it is only a moderate correlation for some factors. However, correlation between two variables does not consider the other influential variables, so partial correlations should be considered. These are shown in Table 3. The first row of the table summarizes the partial correlations described above between angular depth and the various features. All of these correlations are statistically significant ($p<0.05$). The remaining rows show partial correlations of different combinations of parameters. It is noticeable that the angular insertion depth and the cochlear scale are negatively correlated, which is reasonable because, for a particular electrode array and a fixed BID, the bigger the cochlea gets, the smaller the angular insertion depth becomes. We observe strong partial correlation of BID, scale, and electrode length with angular insertion depth, and a weak but significant partial correlation of array diameter with angular insertion depth.

Table 3.

Partial correlation among angular insertion depth, BID, cochlear scale, electrode array length, and electrode array diameter.

	Angular insertion depth	BID	Cochlear scale	Electrode array length	Electrode array diameter
Angular insertion depth	1	0.72 [0.60, 0.81]	$-\mathrm{0.63}[-\mathrm{0.74},-\mathrm{0.48}]$	0.64 [0.49, 0.75]	0.38 [0.18, 0.55]
BID	0.72 [0.60, 0.81]	1	0.59 [0.44, 0.72]	$-0.42[-0.58,-0.22]$	$-0.39[-0.56,-0.20]$
Cochlear scale	$-0.63[-0.74,-0.48]$	0.59 [0.44, 0.72]	1	0.33 [0.12, 0.50]	0.34 [0.14, 0.52]
Electrode array length	0.64 [0.49, 0.75]	$-0.42[-0.58,-0.22]$	0.33 [0.12, 0.50]	1	0.30 [0.10, 0.49]
Electrode array diameter	0.38 [0.18, 0.55]	$-0.39[-0.56,-0.20]$	0.34 [0.14, 0.52]	0.30 [0.10, 0.49]	1

3.6. Angular Insertion Depth Prediction

We constructed a linear regression model for estimating the angular insertion depth. Equation (3) shows the linear equation that we have designed for the angular insertion depth prediction.

$$y=b_0+b_1{x}_1+b_2{x}_2+b_3{x}_3+b_4{x}_4,$$ (3)

Here, $y$, $x_1$, $x_2$, $x_3$, and $x_4$ are angular insertion depth, BID, cochlear scale, electrode array length, and array diameter, respectively. We first built manufacturer-specific models, hypothesizing that manufacturer-specific models may be more accurate than a combined one. The boxplots in Fig. 8 show the prediction errors for different Med-EL electrode arrays when using the leave-one-out models for prediction (average  $R^2=0.71$, $p=5.63e-09$).

Fig. 8.

Prediction error in cases of different MED-EL electrode arrays.

The difference between the actual angle and the predicted angle is considered the angular insertion depth error. Negative values correspond to predictions that are too deep whereas positive values signify predictions that are too shallow. The standard deviation of the overall error is 45 deg. The mean angular insertion depth prediction error for the standard electrode array was higher compared with the Flex28 but lower compared with the Flex24 electrode array. It is worth noting that, because the array diameter term $x_4$ was not found to be significant ($p>0.05$), it was removed from the MED-EL model.

Then, we implemented the regression analysis for the Cochlear CI522 and the HiFocus SlimJ electrode arrays, separately. For the CI522 and SlimJ array models, the electrode array length and diameter are fixed constants; therefore for these models, we excluded the $x_3$ and $x_4$ terms from the model. For the Cochlear CI522 electrode arrays, the model demonstrated a significant relationship ($R^2=0.66$, $p=8.9e-4$) among the considered factors, and the standard deviation of the angular depth prediction error was 29 deg. In contrast, the HiFocus SlimJ electrode array models ($R^2=0.50$, $p=2.94e-5$) resulted in a standard deviation of 40 deg. Figure 9 shows the prediction errors for these two devices. Estimated coefficients for all three of these manufacturer specific models are shown in Table 4. It is worth noting that no values are provided for coefficients corresponding to terms that were excluded as discussed above.

Fig. 9.

Prediction error in cases of CI522 and HiFocus SlimJ electrode arrays.

Table 4.

Estimated coefficients and corresponding $p$ values for different types of electrode arrays.

	MED-EL (Standard, Flex28, and Flex24)		CI522		HiFocus SlimJ
	Estimated value	$p$ value	Estimated value	$p$ value	Estimated value	$p$ value
$b_0$	619.13	0.003	843.65	$4.02e-05$	870.91	$1.294e-06$
$b_1$	32.098	$1.161e-08$	16.083	0.0007	19.222	$3.651e-05$
$b_2$	$-373.2$	$5.839e-07$	$-175.59$	0.0039	$-173.88$	0.002
$b_3$	31.864	$9.367e-07$	—	—	—	—

The estimated coefficients and the $p$ values are provided in Table 4. It is noteworthy that the term $b_o$, which is a residual term that accounts for the variance that is not accounted for by scaling the other parameters, has the largest $p$ value in the MedEl model, which would indicate less statistical confidence compared with the other terms that this value is significantly different from 0.

When the model is constructed using devices from all manufacturers together, a stronger correlation with angular depth is observed ($R^2=0.813$, $p=1.11e-28$) compared with the three individual models.

These results are summarized in boxplots in Fig. 10, which presents the angular depth prediction error based on electrode array length, array diameter, BID, and cochlear scale. The standard deviation of the prediction error of the overall model was 41 deg. For various straight electrode arrays, the model is capable of making reasonable predictions of the angular insertion depth. Model parameters are included in the left column of Table 5. With the overall model, both of the array geometry parameters are significant and are needed to obtain the most accurate predictions.

Fig. 10.

Angular insertion depth prediction error based on electrode array geometry, BID, and cochlear scale.

Table 5.

Estimated coefficients and corresponding $p$ values for the overall model.

	Overall (using cochlear scale)		Overall (using $A$ length)
	Estimated value	$p$ value	Estimated value	$p$ value
$b_0$	529.94	$3.168e-06$	488.76	$1.387e-05$
$b_1$	23.863	$1.942e-14$	23.422	$9.423e-14$
$b_2$	$-252.8$	$2.763e-10$	$-73.503$	$1.494e-09$
$b_3$	17.368	$9.511e-11$	18.033	$4.706e-11$
$b_4$	285.58	0.0004	259.28	0.001

Also shown on the right of Table 5 are parameters for an overall model built using the length $A$ to represent cochlear size instead of our proposed measure of cochlear scale. We used the method proposed by Rivas et al.²⁷ to automatically determine $A$ for each case in our dataset. When using $A$ instead of the cochlear scale, the performance of the model is almost identical. Figure 11 shows the angular depth prediction error based on electrode array length, BID, and cochlear scale. This consideration led to almost similar model performance ($R^2=0.805$, $p=5.75e-28$ and the standard deviation of the prediction error was 41 deg. Schurzig et al.²⁵ presented an angular insertion depth prediction method for MED-EL Flex electrode arrays with a proposed geometric model using basal turn length and the inserted electrode array length. The absolute deviation of the prediction error was $27\mathrm{deg}\pm 18\mathrm{deg}$ for their dataset. We implemented their proposed method and tested it on our MED-EL Flex arrays since this is the type of array that the method was specifically tested with in their paper. We found that it resulted in absolute deviation of the prediction error of $45\mathrm{deg}\pm 37\mathrm{deg}$ on our MED-EL Flex dataset. The increase in the absolute deviation of the error in the case of our dataset might be due to the variabilities present among features in our dataset. In contrast, our prediction model demonstrated a better absolute deviation of $39\mathrm{deg}\pm 25\mathrm{deg}$ on the same cases. According to the Wilcoxon signed-rank test, our model showed statistically significant improvement ($p=0.0018$) in angular insertion depth prediction compared with the geometric model proposed by Schurzig et al.²⁵

Fig. 11.

Angular insertion depth prediction error based on electrode array length, BID, and A-length.

4. Impact on Interventional and Surgical Data Science

Cochlear implantation is an intervention that has yet to significantly benefit from the data science revolution. Surgical techniques are traditionally applied in a one-size-fits-all manner with little patient-specific adaptation to the approach. This has been necessary because the only outcome quality control metrics available to drive innovation have been based on long-term hearing performance after device acclimation, which is at least six months after implantation. The lack of surgical data collection and the long time scale for feedback have made it difficult to strongly link surgical approach to outcomes. However, recent technological developments in automated image processing for CI recipients have enabled collection of a large dataset of accurate postoperative measurements of CI electrode placement for individual patients.^29–31 These data provide new information about surgical approach (the insertion depth of the base of the array) and outcome (the angular depth of the most apical electrode of the array). Analysis of these data has revealed links between electrode placement and hearing outcomes and wide variability in electrode positioning.¹³ Consistent placement of CI electrode arrays is difficult as the size and shape of the cochlea and the various electrode array models are highly variable, and surgeons cannot see inside the cochlea. Using our dataset, we are able to construct a data-driven linear regression model that can provide a preoperative prediction or intraoperative estimation of angular insertion depth of lateral wall CI electrode arrays based on patient-specific cochlear size, the dimensions of the electrode array, and the array base insertion depth, which is directly controllable by the surgeon. The model could be helpful for CI surgeons to make patient-specific preoperative decisions about which electrode array to use and intraoperative decisions about the ideal depth to thread the array into the cochlea to reduce trauma and preserve the residual hearing. With a preoperative CT, the model can be automatically applied with the use of existing methods to automatically segment the cochlea, or it can be used with simple manual measurements of cochlea size. To implement the angular insertion depth prediction model, a surgeon needs to know measurement of the cochlear scale of the patient. Let us assume that the patient has a cochlear scale of 2.85 mm. The cochlear scale can be measured using Eq. (1) after generating the ST mesh from the preoperative CT images implementing automated algorithms.^29,30 In addition, the audiologist has confirmed that the patient has a residual hearing region below 250 Hz, which approximately represents 615 deg inside the cochlea. As our model has an absolute deviation of the prediction error of 32 deg across our entire dataset, the surgeon may choose to target an angular insertion depth of 583 deg, which will provide a 32 deg safety margin, to avoid the insertion of the array into the residual hearing region. If a MED-EL Standard electrode array is used ($\text{length}=31.5\mathrm{mm}$, $\text{diameter}=0.75\mathrm{mm}$), the recommended BID, the signed distance from the most basal electrode to the RW membrane where the array is inserted into the cochlea, can be calculated using Eq. (3) to be 0.51 mm. The left portion of Table 5 provides the corresponding coefficients. Now, if the surgeon wants to use the $A$ length instead of the cochlear scale, he/she needs to replace the cochlear scale value in Eq. (3) with the $A$ length, as measured from a CT scan using manual or available automated tools, and use the coefficients provided on the right portion of Table 5. In this case, if the $A$ length of the patient’s cochlea is 9.30 mm, the recommended BID is 0.65 mm. Using this type of approach, the model can help the surgeon to select the appropriate electrode array and/or make the decision about BID. The use of the model we present could lead to improved hearing outcomes for CI recipients. More generally, this work represents an example of collecting and analyzing the surgical approach and outcome metrics to construct data-driven preoperative and intraoperative prediction models that can improve surgical outcomes. This data-driven approach can provide surgical guidance to reduce the risk of damaging the residual hearing region inside the cochlea.

5. Conclusion

For the proper placement of the electrode array and for the preservation of the low-frequency residual hearing, it is important for a surgeon to know the relative position of the array inside the cochlea. This paper investigates the relationship of the angular insertion depth with the BID, cochlear scale, electrode array length, and array diameter with a dataset composed of multiple commonly used lateral wall arrays [MedEl Standard (5 cases), Flex28 (31 cases), and Flex24 (1 case); Cochlear CI-422 (16 cases); and Advanced Bionics HiFocus SlimJ (33 cases)]. In addition, a model is proposed for the prediction of angular insertion depth based on these features. Our results show how accurately our model can predict the angular insertion depth for straight electrode arrays. Surgeons may use the model to select the best electrode array and/or plan a base insertion depth for a given array that minimizes the likelihood of inserting the array into a region of the cochlea where residual hearing exists.

Careful interpretation of model predictions should include consideration of the potential for positive and negative errors in depth predictions. For hearing preservation subjects, for example, it would be wise to plan a base insertion depth corresponding to a predicted angular depth that is shallower than the depth of preserved hearing structures by some margin, such as 45 deg, to limit the possibility of overshooting into the hearing preservation region. Future work will prospectively evaluate the accuracy of this prediction model.

6. Appendix

Table 6 provides the overall dataset including base insertion depth, cochlear scale, A length, B length, array type and actual and predicted angular insertion depth.

Biographies

Mohammad M. R. Khan received his BSc degree in electrical engineering from Islamic University of Technology, Bangladesh, in 2013 and his MSc degree in electrical engineering from Mississippi State University, Mississippi, in 2018. He is currently a PhD student in the Department of Electrical Engineering and Computer Science, Vanderbilt University. His research interests include medical image processing, statistical modeling, image-guided surgery techniques, and image segmentation.

Robert F. Labadie received his BS degree in mechanical engineering from the University of Notre Dame, Indiana, in 1988. He received his PhD in bioengineering and MD degree from the University of Pittsburgh, Pittsburgh, in 1995 and 1996, respectively. Currently, he is an associate professor in the Department of Otolaryngology/Head and Neck Surgery, Vanderbilt University Medical Center, Tennessee. His clinical specialty is otology with emphasis on surgical rehabilitation of the hearing impaired, including cochlear implantation.

Jack H. Noble received his BE and MS degrees and his /PhD in electrical engineering from Vanderbilt University, Nashville, Tennessee, in 2007, 2008, and 2011, respectively. He is currently a research assistant professor in the Department of Electrical Engineering and Computer Science, Vanderbilt University. His primary research interests include medical image processing, image segmentation, registration, statistical modeling, and image-guided surgery techniques.

Disclosures

Mohammad M. R. Khan and Jack H. Noble declare that they have no conflicts of interest. Robert F. Labadie received consulting fees from Advanced Bionics (a CI manufacturer) and Spiral Therapeutics (a company looking at delivery of drugs and therapeutics to the cochlea in the hopes of improving hearing).