VOR gain calculation methods in video head impulse recordings

2.1Test subjects

The data set from the previous publication [17] was used for all vHIT gain calculations in this manuscript. The vHIT data were collected as part of a prospective cross-sectional multicenter study (OSF Saint Francis Medical Center, Peoria, Ill., USA, and Johns Hopkins Hospital, Baltimore, MD, USA) between August 2011 and December 2012. Twenty-six patients with an acute vestibular syndrome and a mix of underlying disease pathologies [25] were examined. Patient characteristics and diagnoses are summarized in previous publications reporting diagnostic accuracy [17] and type of artifacts [16] in the same population.

2.2Video head impulse test (vHIT) assessment

Fig.1

Figure 1 depicts one clean (Fig. 1A–C, no artifacts) and unclean (Fig. 1D–E, trace with artifacts) vHIT example (eye- and head velocity profile) from one patient with PICA stroke. The unclean vHIT shows trace oscillation artifacts due to intermittent pupil tracking loss. VOR gain has been calculated by (A and D) taking different time points at 40 ms and at 60 ms after HIT onset (‘velocity gain’), (B and E) applying a linear regression (‘regression gain’), or (C and F) comparing the area under the curve for eye (black) and head (grey) velocity (‘position gain’). Note that VOR gain from the same patient and the same vHIT trace resulted in different VOR gains for each calculation method ranging from 0.73–1.1 (traces with no artifact) and from 0.39–1.13 if artifacts changed the morphology of the bell-shaped slow phase curve.

A dataset with 1300 HIT traces in total was used as the basis for our study. While performing the horizontal head impulse test toward each ear, eye and head movements were recorded at the bedside with a vHIT device (ICS Impulse; formerly GN Otometrics, Taastrup, Denmark; now Otometrics division of Natus Medical Inc., Pleasanton, CA, USA). vHIT recordings can be categorized as reflecting either normal or abnormal vestibular function on the basis of VOR gain. A VOR gain close to 1.0 usually reflects a normal value, where as a gain <0.8 is considered abnormal. A gain cutoff of ∼0.7 optimally distinguishes between peripheral and central pathology in patients with the acute vestibular syndrome [17]. Therefore, we classified HIT exams from patients with vestibular neuritis or vestibular strokes into groups with lower (<0.7) or higher gain values (>0.7), since artifacts might have a differential impact on low versus high gain values.

vHITs performed on a diseased ear in patients with clinically-diagnosed vestibular neuritis (i.e., peripheral cases) showed lower gains (i.e., abnormal) on the affected side. The contralateral, clinically-unaffected side showed higher gains within the normal range (though generally below 1.0, presumably because of loss of the contribution from the opposing, affected canal). By contrast, vHITs derived from patients with PICA strokes (i.e., central cases) had higher gains (>0.7, and generally within the normal range) in both ears. PICA stroke was diagnosed based on history, clinical features and diffusion weighted imaging [16]. AICA strokes were excluded because patients might have normal or abnormal HITs with variable VOR gains.

Data from 23 patients, 46 total ears, and 1070 HITs were available for analysis. We classified both clean data (without artifacts, Fig. 1A–C)) and artifactual data (Fig. 1D–F) based on selection criteria reported in a previous publication [16]. Clean records were available for 22 patients, 42 ears, and 539 HITs. We distinguished 6 types of artifacts according to a previously published classification [16]: 1) covert saccades, 2) blinks, 3) trace oscillations, 4) phase shift 5) several peaks, 6) high gain and 7) all artifacts combined (all artifacts pulled together). Traces with infrequent, non-disruptive artifacts (occurring after the head movement) or overt saccades were excluded from the analysis of artifactual data.

2.3vHIT gain calculation methods

Raw vHIT data were extracted from the vHIT device. We applied four different gain calculation methods on the same data using a customized Matlab script (Matlab R2014b, Mathworks, Natick, Mass., USA): (#1a) the ratio between eye and head velocity at 40 ms (Fig. 1A); (#1b) the ratio between eye and head velocity at 60 ms after onset of head movement (Fig. 1A); (#2) the ratio of velocity slopes using regression around peak acceleration (±15 ms) (Fig. 1B) and (#3) the ratio of area under the curves after de-saccading the whole slow phase VOR trace (Fig. 1C).

Calculations #1a, #1b reflect a ratio between velocities at given time points; calculation #2 compares regression slopes between eye- and head velocity around peak head acceleration [3]; and calculation #3 analyzes eye- and head positions before and after the head movement [14]. For calculation #1, the latency of 40 or 60 ms after head movement onset (head velocity exceeding 20deg/s) was used. Calculation #2 is currently not used by commercial devices. For calculation #3 we used the built-in calculation algorithm from Otometrics (Otosuite^® software, v1.2.18) since they use a proprietary method for de-saccading traces and for determining the area under the curve. For simplicity, we call these gain calculation methods here ‘velocity gain’ (specifying either 40 ms [#1a] or 60 ms [#1b]); ‘regression gain’ (#2); and ‘position gain’ (#3) (Fig. 1).

2.4Statistical analysis

Fig.2

VOR gain of HITs with higher gain and lower (abnormal) gain for clean recordings for all four methods. Different letters represent significant differences among the methods inside the normal or abnormal HITs (For variables with the same letter, the difference between these variables is not statistically significant. Likewise, for variables with a different letter, the difference is statistically significant). Means and confidence intervals are model based, due to the nested data.

We recorded 539 HITs from 42 ears from 26 patients. HITs were thus nested within ears (there are several HITs per ear), and ears were nested within patients (data from two ears per patient). The effective sample size were the 26 patients. Due to the nested data, all analyses were fitted with mixed effects models. First, only clean recordings (i.e. without artifacts) were analyzed. VOR gain was fit using calculation method, HIT classification (low/high gain), and the interaction between calculation method and HIT classification as fixed effects, with random intercepts for test nested in ear nested in patient. Position gain was correlated with velocity gain at 60 ms for clean records only, with random intercepts for ear nested in patient. Finally, gain was fit using calculation method, pathology, and artifact, with all possible interactions among the three as fixed effects, and random intercepts for test nested in ear nested in patient. Each artifact was included in a separate model. Multiple comparisons were performed first between the methods for HITs with artifacts, and second between clean and artifact-laden HITs. Multiple comparisons were performed for each artifact and HIT category (low gain/high gain) separately. All analyses were performed in R [22], with the packages nlme [21] for mixed effects models, and emmeans [13] for multiple comparisons.

3.1VOR gain calculations after excluding artifacts (Clean HITs Only)

Fig.3

Correlation between position gain and velocity gain at 60 ms. Grey dots represent the individual HITs, and black dots are mean (±standard error) values per patient. The regression line is derived from the mixed effects model (+/–95% confidence intervals), and takes into account the nested structure of HITs within ears within patients. Note that the regression line shows an offset with an intercept of 0.46. Imperfect calculation algorithms including imperfect de-saccading, removal of negative gain values by the device or data lowpass filtering might lead to a skewed regression line.

Mean VOR gains for clean/filtered higher-gain HITs were all within normal limits, regardless of calculation method, but there were small, statistically significant differences across most calculation methods (range 0.89 to 0.97) (Table 1, Fig. 2). Importantly, these differences were not clinically significant with respect to classifying VOR gain “normality” (i.e., mean results for a given ear, regardless of calculation method, did not cross either the 0.8 [normal vs. abnormal] or 0.7 [central vs. peripheral] thresholds described in the Methods section) (Fig. 2). Similarly, HITs with higher gains were significantly different from abnormal HITs with lower gains, regardless of the calculation method (Table 1).

Mean VOR gains for clean/filtered lower-gain (abnormal) HITs were all below normal limits, and they also showed statistically significant differences across most calculation methods (range 0.45 to 0.63, Table 1, Fig. 2). In addition, there was a strong positive relationship between ‘velocity gain at 60 ms’ and ‘position gain’ for clean HITs (both high and low gain HITs) (conditional r² = 0.77, p > 0.001) (Fig. 3). Table 2 shows the impact of (a) calculation method, (b) “normality” of HITs (high vs. low VOR gain), and (c) the interaction between calculation method and “normality” on mean VOR gain— in a multi-variate regression using a linear, mixed-effects model, all three effects were significant.

3.2Interaction between VOR gain calculations and artifacts (Unclean HITs Only)

Fig.4

VOR gain for clean (white bars) and artifactual (grey bars) higher gain and lower gainHITs for different artifacts and the four methods. Different letters represent significant differences among the methods for gains with artifacts (see legend Fig. 2). Stars denote significant differences between HITs with and without artifacts within a method: *p < 0.05, **p < 0.01, ***p < 0.001. Means and confidence intervals are model-based due to the nested data.

Figure 1D–F shows one example with trace oscillation artifacts (pupil tracking loss) resulting in variable VOR gain values for each calculation method. Gain measures for unclean HITs were more likely to vary by calculation method when gain was high, in the normal range, rather than low, in the abnormal range (Table 3, Fig. 4). HITs with higher gains and artifacts were statistically significantly lower than HITs without artifacts (–0.06 to –0.11, Table 3), but values still remained within normal limits (Fig. 4, Table S1). Abnormal HIT mean gains with artifacts were also generally lower than HITs without artifacts (–0.07 to +0.01, Table 3) but not statistically significantly so when averaged across artifact types (Table 3, “all artifacts” column).

3.3Impact of specific artifacts

The influence of each artifact type on mean VOR gain is shown in Tables 3, 4, and 5 and mean VOR gains with confidence intervals are shown in Fig. 4.

Covert saccades were present in some HITs. The presence of such saccades presumably reflected incomplete de-saccading by the vHIT software, but they did not affect VOR gains meaningfully for either high or low VOR gains in any calculation method (Table 3).

Blinks caused intrusive trace oscillations and a clinically and statistically significant mean gain reduction across all calculation methods in HITs with higher gain (range –0.51 to –0.29) (Table 3). Mean normal (higher) VOR gains when blinks were present fell into the abnormal (lower) range (mean gain range 0.41 to 0.62) (Table S1). Abnormal HIT gains were not significantly affected by eye blinks except when the velocity gain at 40 ms calculation was used (Table 3).

Trace oscillations (pupil tracking loss) showed small gain reductions in higher-gain HITs (range –0.06 to –0.12) that were statistically significant for three calculation methods (Table 3), but calculated gains still remained within normal limits (range 0.81 to 0.91, Table S1). Trace oscillations reduced VOR gains by small amounts for lower-gain (abnormal) HITs as well (range –0.01 to –0.10), statistically significant for two of the four calculation methods (Table 3). Again, all gains from lower-gain HITs remained in the abnormal range (Table S2).

Phase shift between head and eye velocity traces showed a strong interaction between artifact and calculation method (Table 3). For HITs with higher gains, there were variable differences (range –0.13 to +0.19) with regression gain showing a statistically significant increase. For lower-gain (abnormal) HITs, there were variable differences (range –0.19 to +0.09) with velocity gain showing a statistically significant decrease. There was, however, no impact on the classification of gain as normal vs. abnormal or central vs. peripheral for any calculation method for either cutoff (0.7 or 0.8).

More than two peaks in the VOR slow phase (no bell-shaped curve, often an iatrogenic artifact) caused high-gain HITs to be lower (range –0.03 to –0.11), statistically significant in three of four calculation methods. However, none of these reductions put results into the abnormal range (Table S1). By contrast, lower-gain (abnormal) HITs were statistically-significantly higher for two calculation methods (position +0.08 and velocity +0.12) (Table 3), and these pushed abnormal gains towards the normal range (Table S2).

High gains (likely due to goggle slippage or false calibration) caused a clinically-meaningful and statistically significant increase in higher-gain (normal) HITs (range +0.21 to +0.55). Resulting mean gains were above normal (range 1.1 to 1.43) (Table S1). High gain artifacts in abnormal HITs still showed abnormal values in the “peripheral” cause range (gain < 0.7) except when using the regression gain calculation which was highly susceptible to this artifact (unclean mean gain 1.22 vs. clean mean gain 0.37, p = 0.001).

4.1Limitations

Our study involved many head impulses, but from only 26 subjects with a specific clinical presentation (acute vestibular syndrome), so results may not generalize to other patient populations. Although we assessed the most commonly used calculation methods (including those employed by US FDA-approved, production-line vHIT systems currently being sold in the US and Europe), other methods were not tested in our study (e.g., point by point gain average over different time windows using whole or partial velocity profile, ascending or descending window, ±12 ms around peak head velocity).

Although there was a strong correlation between velocity and position gain, with found a skewed regression line (Fig. 3): At lower gains, calculated position gain was rather too high or – vice versa – velocity gain could have been too low. At higher gains velocity gain was higher than position gain. This phenomenon might be due to imperfect gain calculations including imperfect de-saccading algorithms, removal of negative gain values by the device or data low pass filtering. We had, however, no gold standard (scleral search coil) measures to determine the “true” gain for any of our traces.We used raw data from only a single brand of vHIT device in order to obtain uniform data, potentially limiting the generalizability of our results; however, lateral vHIT gains computed with the same technique appear to be similar across vHIT systems [5]. Finally, the data were limited to the lateral impulses only. HITs in the RALP or LARP plane seem to be more prone to artifacts and these artifacts might be more difficult to classify.