The experiments were approved by the Johns Hopkins institutional review board and informed consent was obtained from all the participants. Twelve healthy volunteers (mean age 29 years, 9 females) participated in this study. SVV and ocular torsion were recorded simultaneously in a completely dark room. The head was immobilized using a molded bite bar while the visual stimulus appeared on a CRT monitor (1280 pixel by 1024 pixel) 135 cm away from the subject. We mounted the bite bar on a rotary motor (Zaber Technologies Inc., Vancouver, BC, Canada) in order to change the head tilt position remotely. Each subject participated in two experiment sessions, one with the head tilted 20° to the right, and one with the head tilted 20° to the left, in a random order across subjects. We chose to tilt the head on the body as it has been previously shown that it can produce a relatively large SVV tilt and aftereffect (Wade, 1968). In each session, the subject stayed on the bite bar for the whole time. The SVV was first recorded in upright position, and after 100 trials the bite bar was tilted remotely to record 500 trials while the head remained in the static tilt position. The bite bar was then tilted back to upright to record 150 more trials (total of 750 trials). We added a 30-s pause in the SVV paradigm after each time the head changed position (Figure 1D), in order to avoid residual effects from the semicircular canal stimulation during head movement.

We used RealEyes xDVR system manufactured by Micromedical Technologies Inc., Chatham, IL, USA and a custom software to record torsional eye position. This system uses two cameras (Firefly MV, PointGrey Research Inc., Richmond, BC, Canada) mounted on goggles to capture infrared images of each eye. To measure and track torsional eye position, we used a method developed by our group (Figure 1A) that operates binocularly in real time at 100 Hz and with a noise level less than 0.1° (Otero-Millan et al., 2015).

The software that controlled the SVV paradigm was written in Matlab (Mathworks) using Psychtoolbox (Kleiner et al., 2007). A red line (length: 7.6° of visual angle, width: 0.13°) was presented in random angles around a fixed, red dot at the eye level (diameter: 0.33°; Figure 1B). The perceived line orientation was measured using a two alternative forced choice (2AFC) task. In each trial, subjects clicked the right or left button on a controller to report whether the line was tilted to the right or left of what they perceived as upright. The line orientation was randomly selected within a range that was adjusted in blocks of 10 trials (Figures 1B,C). In each block, five different angle orientations were presented in the top of the visual field (always radiating from the fixation point) and equivalent five angles were presented in the bottom of the visual field. In each trial, the fixation dot appeared first, and then after 1 s, the visual line was presented for a minimum of 300 ms and maximum of 1.5 s until subject responded. If the subject did not respond within 1.5 s, the line disappeared and a new trial started after a button click. In such cases, the angle that was missed was presented again at a later time within the same block to ensure that all angles were presented exactly once. At the beginning of the paradigm, the angles were selected randomly from the entire range of 360°, but as the paradigm continued the range of probing angles was adjusted frequently to precisely track temporal changes in subject’s perception. To make this adjustment, every 10 trials a new range was calculated that: (1) was centered around the SVV calculated from the responses in previous 30 trials (for SVV calculation see “Data Analysis” Section); and (2) had an amplitude that decreased by half until the paradigm reached the 9th block, after which it remained constant at 10° for the rest of the trials (Figures 1C,D). Thereby, the SVV paradigm could adapt, track changes in subject’s perception, and it was not biased by making any prior assumption about the SVV value. Every time the head changed position from upright to tilt or tilt to upright, the paradigm reset to the starting range of 360° (Figure 1D).

SVV was calculated by fitting a psychometric curve to the responses using a logistic function and a generalized linear regression model (Matlab fitglm). The SVV value was the angle at which the probability of left or right responses was 50% (point of subjective equality). The SVV precision was calculated as the difference between the 50% and 75% points on the psychometric curve.

In order to compare ocular torsion and SVV responses, we first calculated the average torsional position of the two eyes during each trial in the SVV paradigm. Then within a window of 100 trials, we calculated the average ocular torsion and fitted a psychometric curve to the responses from these 100 trials to calculate the SVV value (Figure 1E). The window then advanced in steps of 50 trials to obtain more SVV and ocular torsion values. The first 50 trials were discarded as the range of angles in these initial trials was not narrow enough to get a reliable SVV value (Figure 1D). We used a simple linear regression to measure the drift over time and estimate the rate of change for both SVV and ocular torsion. To calculate correlations across subjects, we first averaged the values for the right and left head tilts and then used Spearman method to obtain the coefficient. For comparisons, we used t-test with a significant p-value less than 0.01.

Our results show that perception of upright may drift over time during a static head tilt, reflected by a significant change in SVV responses in the direction of the head tilt. This finding is consistent with previous reports of SVV drift during head tilt (Wade, 1970; Tarnutzer et al., 2013; Kheradmand et al., 2015). Ocular torsion is a source of error in upright perception as the orientation of the images on the retina changes with ocular counterroll (Wade and Curthoys, 1997; Pavlou et al., 2003; De Vrijer et al., 2009). Here we addressed a possible link between the drift in upright perception and torsional eye position by tracking SVV responses and ocular torsion simultaneously. We found no significant correlation between the drifts in upright perception and ocular torsion; that is, subjects with larger drift in ocular torsion did not necessarily have larger drifts in SVV responses. In addition, while the average torsion drift was in the same direction as the SVV drift for the left and right head tilts, the drift in torsion was smaller and more consistent across subjects than the SVV drift, which varied from subject to subject (Figure 3). Therefore, the torsional eye position—or its driving input from the otoliths—may not be the source of drift in upright perception during head tilt. A similar drift has been shown when upright perception was measured by a haptic task, also suggesting that the drift cannot be entirely related to visual errors induced by ocular torsion (Wade and Day, 1968a; Tarnutzer et al., 2013).

We found a significant SVV bias in the tilt direction after the head returned to upright position. There was however no significant aftereffect in ocular torsion following head tilt. Others studies have also shown aftereffect in upright perception using either visual (i.e., SVV) or haptic tasks, however ocular torsion has not been measured previously (Wade and Day, 1968a,b; Tarnutzer et al., 2013). With the visual task, the magnitude and direction of aftereffect were not different in the supine and upright body positions, suggesting that neck proprioception was the primary source of adaption in upright responses during head tilt (Day and Wade, 1968).

We did not find a significant correlation between the drift in SVV responses and aftereffect consistent with findings from Tarnutzer et al. (2013). However, while most of their subjects showed a drift in SVV responses during head tilt, there was no significant drift in the group data due to variability among individual subjects. In addition, some subjects showed aftereffects in the opposite direction of the drift while in our results the SVV aftereffect and drift were consistently in the same direction. The difference in these findings could be related to how the head was tilted in each study. In our experiment the head was tilted on the body whereas in Tarnutzer et al. (2013) the head and body were tilted en bloc. As previously reported, SVV aftereffect with head tilt alone could be larger than the aftereffect with whole-body tilt (Wade, 1970). Also, when SVV was measured in different head and body tilt positions, the error was aligned with the head axis as the reference frame, and not the trunk axis (Guerraz et al., 1998; Tarnutzer et al., 2010). These findings together suggest that the adaptive changes in upright perception are mainly related to the head position relative to the body, and not the trunk or head position relative to gravity.

An intuitive and effective computational model for static upright perception in the dark is based on the Bayesian approach (MacNeilage et al., 2007; De Vrijer et al., 2009; Clemens et al., 2011). In this framework, the upright estimate is determined by a weighted average of existing knowledge (prior) and noisy sensory information. Here we examined whether the drift and aftereffect pattern in SVV responses could be simulated by changes in individual parameters of the Bayesian model (De Vrijer et al., 2009).

The SVV error in the Bayesian spatial-perception model is determined by the head-in-space and eye-in-space estimates:

In Equation (1) μ_svv represents SVV error, H_s actual head-in-space position, μHS˜ estimated head-in-space position, E_H actual eye-in-head position, and μEH˜ estimated eye-in-head position. As a convention, all the sensory inputs are denoted by the hat symbol (∧) and all the estimates obtained by sensory integration are denoted by the tilde symbol (~). For example, HS^ represents the head orientation in space as measured by the head-in-space sensors, and HS˜ represents the final head-in-space estimate by the brain. Among the sensory inputs to the model, the head-in-space input (HS^) is noisy but unbiased by systematic errors (with a variance of σHS^2), and the prior (with variance of σHSp2) is taken into account to estimate head position (μHS˜). Since we spend most of our time in upright position, the prior for head position is a Gaussian distribution that peaks at zero (i.e., upright position). This results in an underestimation of upright at larger head tilt angles as the weighted estimate of head position is biased by the prior (i.e., A effect). De Vrijer et al. (2009) added a free parameter to the model (ΔE_H) that accounts for an error in estimating ocular torsion by the brain. This “uncompensated” ocular torsion can explain the SVV error in the opposite direction at smaller head tilt angles (i.e., E effect). The model also allows for a lower precision of the head-in-space inputs (σHS^2) at larger head tilt angles. This is formulated by

in which α₀ represents the noise in upright position and α₁ accounts for the proportional increase of noise with head tilt. Note that this model assumes a vertical orientation of the trunk. Thus, the sensed head orientation in space (HS^) is a combination of the otolith and proprioception signals.

We used individual fit parameters of the Bayesian model to simulate the pattern of SVV drift and aftereffect. These parameters included the prior for head position (H_Sp), the head-in-space sensory input (HS^) and the sensory noise with tilt angle (α₁; Figure 5). Since the model does not allow for large changes in SVV from small changes in ocular torsion—see Equation (1)—and because we did not observe a large drift in ocular torsion, ΔE_H was not included as a possible source of SVV drift. Equation (1) was also modified to allow the head position prior (H_Sp) to drift as a function of time with values different from zero:

Among all the fit parameters, only the drift in head-in-space sensory inputs (HS^) could produce the pattern of SVV drift and aftereffect in our data (Figure 5A). The head-in-space information is a combination of the otolith and proprioceptive inputs. Since the changes in ocular torsion during head tilt were much smaller than the changes in SVV and there was no significant correlation between the drifts of SVV and ocular torsion, the proprioceptive inputs—and not the otoliths—could be the main source of the drift in the head-in-space sensory input (HS^). This finding is line with previous studies that suggest neck proprioception is the source of SVV drift and aftereffect (Wade and Day, 1968a; Wade, 1970). Other simulation scenarios included the drift of head prior (H_Sp) toward actual tilt position and the increase in sensory noise (α₁) during head tilt. In case of the drift in H_Sp, we examined whether a shift of the prior toward the head tilt position could explain the drift in upright perception. In such case, however, the SVV would drift in the opposite—and not toward—the direction of the head tilt (Figure 5B). In case of the increase in α₁, we examined whether the drift in upright perception could be related to an increase in sensory noise during head tilt. In this case, HS^ becomes less reliable over time during head tilt, which gradually increases the weight of the head prior in SVV responses. This scenario also produces an SVV drift toward the actual head position. However, when the head returns to upright, the sensory noise would reduce back to the baseline (α₀) and thus it does not result in an aftereffect (Figure 5C). While the drift in head sensory inputs (Figure 5A) can explain our result, it does not account for a lack of correlation between the magnitudes of SVV drift and aftereffect. Thus, it is possible that other scenarios above (Figures 5B,C) also affect the SVV drift and cause variability among individuals while the aftereffect is not influenced. Future studies will have to consider more sophisticated state-space models such as Kalman filters to address the possibility of SVV drift based on the internal estimates of the model and not just its parameters.

In sum, our results show that SVV drift during head tilt and its corresponding aftereffect in upright position were not correlated with changes in ocular torsion. These results along with simulations from the Bayesian spatial-perception model suggest that proprioception could be the source of drift in upright perception during head tilt, and subsequently the aftereffect in upright position. To verify this hypothesis, SVV drift and ocular torsion could be measured under different combinations of whole-body and head-on-body tilts. For example, we expect to find smaller drifts in upright perception during whole-body tilts. To estimate head orientation from neck proprioception, the brain must also estimate body orientation (Clemens et al., 2011; Fraser et al., 2015). Future experiments could dissociate the effects of neck and body proprioception and also test if the results are consistent for small and large tilt angles.