An Implanted Vestibular Prosthesis Improves Spatial Orientation in Animals with Severe Vestibular Damage

Experimental design.

Two female rhesus monkeys (M1 and M2) were trained to perform a psychophysical task that assayed their perception of head orientation relative to gravity, using a testing paradigm (Lewis et al., 2008, 2013) derived from the subjective visual vertical (SVV) task commonly used in human subjects (Dieterich and Brandt, 2019; Cho et al., 2020). Each monkey was trained to rotate a light bar about the roll (naso-occipital) axis using a small steering wheel with the goal of aligning the light bar parallel to gravity. Once learned, each monkey performed this task in the dark without any visual orientation cues while being dynamically tilted (sinusoidal or pseudorandom trajectories) about an earth-horizontal roll axis centered between the ears with the head immobilized related to the motion platform. Eye movements generated by angular and linear motion trajectories were also measured (in complete darkness) to quantify the efficacy of the peripheral canal, VI, and otolith inputs.

SVV testing and eye movement recordings were performed in both monkeys in four vestibular states. First, the normal monkeys were tested, then bilateral vestibular hypofunction (BVH) was induced by destroying one ear with a labyrinthectomy and exposing the ear implanted with the three VI ampullary electrodes to aminoglycosides; and then the animals were tested in this BVH state. The VI was then activated and provided both a tonic (resting) rate of electrical stimulation to the three ampullary nerves of the instrumented ear and motion-modulated stimulation that was based on the angular velocity of the head in the plane of each of the three canals (chronic VI stimulation). During the months of VI stimulation, the monkeys habitually experienced motion-modulated stimulation while moving freely in their housing and were intermittently tested on the SVV task with the modulation-OFF (tonic stimulation only) and modulation-ON (tonic and motion-modulated stimulation), using the transfer function relating angular head velocity to electrical stimulation that was habitually experienced. M2 was also tested with a modulation-ON ×2 state, which used a transfer function with twice the sensitivity of the chronically experienced stimulation.

SVV training and testing.

To quantify perceived head orientation relative to gravity, we used an SVV task similar to the version we previously used (Lewis et al., 2008, 2013). Briefly, the monkey sat in a primate chair with the head immobilized and rotated a circular steering wheel mounted inside the chair. An encoder measured the wheel position and rotated the roll orientation of a light bar (0.20 m long), with 1° of wheel rotation causing 1° of bar rotation. The light bar was displayed on a monitor 0.55 m in front of the animal and was viewed through a circular aperture. The room and monitor were otherwise black, and the center of the monitor, aperture, and light bar were aligned with the center of the head of the animal. During training, the monkeys learned to align the light bar parallel to a reference line, which was aligned with gravity. After this was mastered, the reference line was removed, and the monkeys learned to align the light bar parallel to gravity without visual cues. During training they were tilted in roll and accurate responses were required to receive a reward. Adequate training required ∼6 months, and a brief training session with the reference bar on preceded each testing session to reinforce the learned behavior.

The testing paradigm appeared identical to the training task to the animal but did not require accurate responses to receive a reward and provided no feedback about response accuracy. Using this test, we previously recapitulated in rhesus monkeys many SVV features observed in humans, such as A and E effects and tilt illusions during centrifugation (Merfeld et al., 2001; Lewis et al., 2008; Fraser et al., 2015). While the task used in our prior studies was intermittent (the light bar was extinguished after each trial and reappeared at a random orientation), the current study used a continuous SVV task (the light bar remained visible throughout the trial) to allow recordings that were uninterrupted. This task was challenging for the animals because the bar position was continuously perturbed, which required the animal to continuously rotate the wheel to cancel these perturbations. The perturbations had two components: colored noise, which was Gaussian noise with an SD of 3.0° that was filtered using a first-order, low-pass filter with a cutoff of 0.03 Hz; and an instability that caused the bar to have an angular acceleration proportional to its current velocity. Update and sampling rates were 60 Hz.

Motion protocols and data analysis for SVV and VOR testing.

For SVV testing, the primate chair was mounted on a platform rotated by a computer-controlled servomotor. The platform tilted in roll about an earth-horizontal roll (naso-occipital) axis passing through the center of the head between the ears, motion that modulates activity in the four vertical canals and the four otolith organs (primarily, the two utricles for the relatively small tilt angles we used). Sinusoidal trajectories had a fixed amplitude of ±20° and were performed at frequencies of 0.005, 0.01, 0.05, 0.1, 0.2, 0.3, and 0.4 Hz. The order of testing was randomized with regard to frequency. The pseudorandom trajectory was a sum of sines profile that included three sinusoids at different frequencies (0.051, 0.13, and 0.2 Hz). Chair position was measured by an encoder, sampled by LabVIEW, and stored by the computer controlling the visual display. For each SVV trial, a trained operator reviewed the light bar and chair tilt angle data to identify segments in which the monkeys were not attending to the task—based on extensive observation during the training and testing trials, these periods of inattention were defined as data segments during which the animal did not turn the wheel for >10 s and/or data segments where the light bar deviated >45° from upright, and these data were excluded from further analysis. The “perceived roll tilt angle” was calculated as the difference of the head/chair tilt angle and the SVV error (angular distance of the light bar from upright; Lewis et al., 2008). The tilt gain was determined for each frequency and vestibular state and was defined as the ratio of the [amplitude of the perceptual roll tilt report]/[amplitude of the actual tilt].

In the normal animal, roll tilt modulates activity in the four vertical canals and the four otolith organs (primarily the two utricles for the small tilt angles we used, which are orientated close to the earth-horizontal plane with the head upright). We therefore measured VOR responses during roll tilt and during interaural (IA) linear translation (to measure utricular function) in each vestibular state. Specifically, we measured the angular VOR during roll tilt about an earth-horizontal axis using sinusoidal rotations (0.5 Hz, 20°/s peak velocity), and the IA linear VOR along an earth-horizontal axis during sinusoidal translations (0.7 Hz, 44 cm/s peak velocity). Three-dimensional eye movements were recorded in one eye using standard search coil methods and were calibrated with a gimbal, with 5 V equaling 20° for horizontal and vertical eye movements and 15° for torsional eye movements. Eye movement traces were desaccaded using standard methods (Lewis et al., 2010), and each trace was manually reviewed by an experienced user.

Instrumentation.

The following four surgical procedures were performed under general anesthesia. (1) A head bolt was attached to the skull using inverted screws set in dental cement. Relative to the stereotactic zero, it was pitched forward ∼18° so the lateral and vertical canals were close to alignment with earth horizontal and vertical, respectively, when the head was immobilized (Blanks et al., 1985). A headcap affixed to the bolt housed the sensors of the VI and stimulation electronics and connectors for the ear electrodes and eye coils. (2) Frontal and sagittal eye coils (Judge et al., 1980) were implanted in one eye under the conjunctiva. (3) Three multisite, platinum stimulation electrodes (MED-EL) were inserted into the three canals of one ear. After drilling openings into the lateral, posterior, and superior semicircular canals near the ampullae, the electrodes were inserted and their position was adjusted using a micromanipulator to maximize eye movements (generated by short pulse trains) in the direction corresponding to the plane of the stimulated canal while simultaneously minimizing eye movements corresponding to other canals or facial movements (for surgical details, see Merfeld et al., 2007; Lewis et al., 2013). After optimally positioning each electrode, the leads were sutured to the skull and then tunneled subcutaneously to the headcap. (4) After confirming postoperatively that the electrodes functioned adequately, a labyrinthectomy of the contralateral ear was performed.

Vestibular ablation.

We intended to use aminoglycosides to ablate vestibular function and began with intratympanic (IT) injections into each ear using gentamicin (26.7 mg/ml; Syed et al., 2015), evaluating the effect of ablation by measuring angular and linear VOR responses. Rhesus monkeys appeared to be fairly resistant to this mode of ablation, however, as 10 bilateral IT injections in M1 only modestly reduced the VOR. We next moved to systemic aminoglycosides using intramuscular streptomycin at doses described to be ototoxic in monkeys (Igarashi et al., 1966). After two courses of streptomycin administered intramuscularly in M1, substantial VOR responses remained, although they were diminished by ∼40% compared with normal. To finish the ablation process in M1 (and the complete ablation process in M2, where courses of IT and intramuscular aminoglycosides were omitted), we used direct irrigation of the endolymph in the implanted canal with gentamicin solution during electrode surgery. After verifying adequate function in the implanted electrodes postoperatively, in a separate surgery the contralateral ear was destroyed with a labyrinthectomy. The outcome of this process was a complete loss of vestibular function on the nonimplanted side and severe damage in the implanted ear, which resulted in the angular and linear VOR becoming undetectable. While the damage incurred in the implanted ear was severe enough to eliminate measurable VOR responses, the persistence of minimal residual function in the implanted ear could not be excluded, but this possibility would not affect the results of the study.

Tuning of the canal electrodes.

Electrode tuning was performed as previously described (Merfeld et al., 2007; Lewis et al., 2010). Briefly, eye movements were recorded while pulse trains were applied using one of the three canal electrodes. Like the implant (below), pulses used in tuning experiments were biphasic and charge balanced, with each pulse 200 μs in duration and the cathodic and anodic pulses separated by 25 μs without current. During electrode tuning, all of the different current and return sites for each electrode (30 potential combinations) were tested, using different current amplitudes for the pulses. The goal was to optimize the size of the eye movements in the plane of the canal while concurrently minimizing orthogonal eye movement responses and facial nerve activation. The electrodes were also tested simultaneously with phase-locked pulses to confirm that eye movements were a linear superposition of those produced by stimulating each canal. After electrode tuning, we found that phase-locked activation of the anterior and posterior canal electrodes (simulating the VI stimulation produced by roll tilt of the head) produced substantial in-plane torsional eye movements but very small out-of-plane vertical and horizontal eye movements (each of which was <10% the size of the in-plane torsional eye movement response). The current pulses for each canal electrode were thereby set at 240 μA (anterior), 140 μA (posterior), and 300 μA (lateral) for M1; and 120 μA (anterior), 180 μA (posterior), and 100 μA (lateral) for M2.

Implant stimulation.

Angular head velocity was transduced by a three-dimensional piezoelectric angular rate sensor fixed in the head cap and positioned so that the three sensor axes aligned with the sensitive axes of the three canals, and was filtered with a first-order high-pass filter with a time constant of 10 s (Lewis et al., 2010). Like our earlier VI experiments studying the VOR, the transfer function relating filtered angular velocity to pulse-rate modulation was a hyperbolic tangent that was nearly linear over most of the physiologic range of head velocities but was saturated at high and low velocities. The linear component of the transfer function for both monkeys had a slope of 3 pulses/s (pps)/(°/s). The motion-modulated component of the VI stimulation was superimposed on a tonic stimulation rate of 200 pps, which was well above the normal baseline firing rate of vestibular afferents (∼100 spikes/s; Goldberg and Fernández, 1975) so it could be modulated upward or downward for ipsilateral or contralateral head rotations (Merfeld et al., 2007; Lewis et al., 2010). Once the VI was activated, this chronic motion-modulated stimulation occurred continuously (e.g., while housed) except during experiments. During experimental sessions, the VI was set to modulation-OFF and modulation-ON, and for M2, modulation-ON, with the sensitivity doubled to 6 pps/(°/s) (modulation-ON ×2), with the order of testing randomized.

Characterizing stimulation of the vestibular end-organs.

As described above (concerning electrode tuning), it is likely that some current spread beyond the targeted canal ampullary nerves to affect other canal or otolith afferents. Of particular interest in this regard is the potential effect that utricular nerve stimulation could have on tilt percepts, since the utricles respond directly to gravitational (and linear acceleration) stimuli. Current spread to utricular afferents cannot be excluded in our study—single-unit recordings would be required to positively identify VI effects on utricular nerve activity, and even if they were not identified in the recorded units, modulation of a subset of utricular afferents still could not be completely excluded. For these reasons, the potential behavioral effects of utricular stimulation deserve consideration with regard to our experimental design. We summarize these behavioral effects here and explain why any potential utricular nerve stimulation should not have a substantial effect on our experimental results. (1) Unlike our prior VI perceptual study (Lewis et al., 2013), which involved tonic electrical stimulation of a vertical canal with the animal stationary, the present study examines motion-modulated stimulation during roll tilt. The important issue introduced when stimulation is time variant is the temporal relationship between actual head tilt and the perception of head tilt. Specifically, the VI sensors (like the canals over the physiologic range of head rotations) transduce angular head velocity, and the VI stimulation targets the ampullary nerve, whose afferents normally encode head velocity. If tilt percepts were because of canal stimulation, as we propose, they would be appropriately in phase with head motion during sinusoidal motion (e.g., tilt percepts would be maximum when head tilt is maximum, although head velocity is transiently zero). In contrast, the utricles sense head position relative to gravity, and their afferent firing therefore encodes angular head position, not head velocity. If tilt percepts during VI stimulation were substantially because of utricular nerve stimulation; therefore, tilt percepts would be proportional to head angular velocity. During sinusoidal motion, tilt percepts would peak when head angular velocity peaks, which is 90° out of phase with the veridical response (e.g., tilt percepts would be absent when head tilt is maximum because angular velocity is zero). Our results (see below) show that tilt percepts were approximately in phase with head tilt position (not head velocity), which is inconsistent with a substantial perceptual contribution derived from utricular nerve stimulation. (2) There are no experiments that have examined tilt perception during isolated electrical stimulation or ablation of the utricular nerve—as described below, eye movements produced by utricular nerve stimulation have been recorded (Suzuki et al., 1969), but not tilt percepts. Furthermore, SVV deviations toward the damaged labyrinth or vestibular nerve have been widely described in human patients (Min et al., 2007), but none of these reports identify the effects of isolated utricular damage on tilt perception. The closest analogs in the literature are the tilt percepts produced by surface galvanic electrical stimulation in human subjects, and as discussed in detail in the study by Fitzpatrick and Day (2004) and in our prior study (Lewis et al., 2013), depolarization or hyperpolarization of the entire utricular macula or afferent nerve should not induce substantial tilt percepts, based on the anatomy of hair cell orientation in the macula. (3) Prior work using binocular eye movement recordings (Suzuki et al., 1969) showed that electrical stimulation targeting the utricular nerve produced a static torsional and vertical offset between the two eyes. We specifically examined this issue in our 2013 study (Fig. 1 in Lewis et al., 2013) using binocular eye coils and found no evidence that such a torsional or vertical offset occurred during tonic vertical canal stimulation. While we cannot recapitulate these findings in our current study since eye recordings were monocular, the data presented in our 2013 study (Lewis et al., 2013) provide strong evidence that stimulation targeting only the vertical canals is capable of inducing tilt percepts. Together, these behavioral observations provide considerable evidence that tilt percepts in our experimental preparation were primarily because of canal activation, and that while utricular afferents may have been stimulated, their contribution to tilt perception was probably small.

Experimental protocol.

In each vestibular state (normal, BVH, VI stimulation), the animals were typically tested three times/week. Because of delays associated with vestibular ablation and surgical recoveries, testing in the different states was separated by substantial periods of time. For M1, the experimental schedule was as follows: (1) 33 testing sessions in the normal state over a period of 11 weeks; (2) after a delay of 142 weeks (because of repeated aminoglycoside administrations), the VI was implanted and tuned; (3) 17 weeks later, the contralateral ear was ablated; (4) 18 weeks later, testing in the BVH state began and included 28 testing sessions performed over 14 weeks; and (5) 34 weeks later, the VI was activated, and 4 d after activation M1 was tested 11 times over 14 weeks in the VI-stimulated state. For M2, the experimental schedule was as follows: (1) 21 testing sessions in the normal state over a period of 8 weeks; (2) after a delay of 23 weeks, the VI was implanted and tuned; (3) 5 weeks later, the contralateral ear was ablated; (4) 13 weeks later, testing in the BVH state began and included 21 testing sessions performed over 18 weeks; and (5) 48 weeks later, the VI was activated, and 4 d after activation M2 was tested 37 times over 23 weeks in the VI-stimulated state.

Statistical analysis.

For single-frequency trials, a gain and a bias (i.e., a consistent offset) relating bar angle to chair angle were determined using a least-squares linear regression for each subject and frequency. Specifically, the following model: θ¯̂f,s,m=-Gf,s,m·θ¯f,s,m + Bf,s,m,(1) was solved to determine a gain parameter Gf,s,m and a bias parameter Bf,s,m, where θ¯f,s,m is the chair tilt angle for frequency f, state s (normal, ablated, modulation-OFF, modulation-ON), and monkey m, and θ¯̂f,s,m is the corresponding perception of tilt angle. Each regression also yielded an SE for each parameter, and these were adjusted to correct for autocorrelation (hac function, Econometrics Toolbox, MATLAB 2017), since ordinary least-squares regression can underestimate SEs when data are temporally correlated. This resulted in SEs eG,f,s,m and eB,f,s,m for Gf,s,m and Bf,s,m, respectively, which were used to compute statistical significances pG,f,s,mand pB,f,s,m. Sum of sines data were also analyzed using regressions; three gain parameters and one bias parameter were determined, which related bar angle to three sinusoids at the stimulus frequencies. Specifically, the following model: θ¯̂sos,s,m=-(Gsos1,s,m·6°·sin(2π·0.051t-) + Gsos2,s,m·5°·sin(2π·0.13t-)+Gsos3,s,m·6°·sin(2π·0.2t-)) + Bsos,s,m,(2) was solved to determine gains Gsos1,s,m, Gsos2,s,m and Gsos3,s,m, and bias Bsos,s,m, where θ¯̂sos,s,m is the perception of tilt angle during sum of sines motion for state s and monkey m and t- was the time vector corresponding to θ¯̂sos,s,m. As described above, SEs and p values were determined for each of the four parameters. Determining phase by adding additional terms (i.e., chair position lagged by 90° for single sinusoids; cosine terms for sum of sines) did not improve fits. This was expected because the bar angle was generally in phase with chair tilt.

To compare gains for normal versus ablated, a linear regression was performed on data pooled across both states, using the model: θ¯̂f,s=normal,mθ¯̂f,s=ablated,m=-Gf,s=normal,m·θ¯f,s=normal,m + Bf,s=normal,m-Gf,s=ablated,m·θ¯f,s=ablated,m + Bf,s=ablated,m=-Gf,s=normal,m·θ¯f,s=normal,m + Bf,s=normal,m-Gf,s=normal,m + ΔGf,s=ablated,m·θ¯f,s=ablated,m + Bf,s=normal,m + ΔBf,s=ablated,m=-Gf,s=normal,m·θ¯f,s=normal,mθ¯f,s=ablated,m + ΔGf,s=ablated,m·0kx1θ¯f,s=ablated,m+Bf,s=normal,m·1kx11lx1 + ΔBf,s=ablated,m·0kx11lx1,(3) which was solved to determine gain Gf,s=normal,m, ΔGf,s=ablated,m, the difference between gains for normal and ablated, and bias parameters Bf,s=normal,m and ΔBf,s=ablated,m, where k is the length of vector θ¯f,s=normal,m and l is the length of vector θ¯f,s=ablated,m. SEs eG,f,s,m, eΔG,f,s,m, eB,f,s,m, and eΔB,f,s,m, and corresponding statistical significances pG,f,s,m, pΔG,f,s,m, pB,f,s,m, and pΔB,f,s,m were determined for each of the four parameters after performing the corrections described above. This approach yielded the same gain values as above, but provided the SEs and statistical significances for the differences in gains. The same approach was used to compare modulation-OFF and modulation-ON, as well as for sum of sines data. The 95% confidence intervals for these analyses are seen in Figures 4 and 5. When comparing states across multiple frequencies, Fisher's method (i.e., Fisher's combined probability test; Fisher, 1925; Mosteller and Fisher, 1948) was used to calculate a p value that combined statistical significances for individual frequencies, pΔG,f,s,m. Comparisons between animals and between single-sinusoid and sum of sines results were made using the nonparametric Wilcoxon signed-rank test.