Using the probability of recovery from an acute unilateral vestibulopathy to determine the severity of the vestibular loss: Development of a multivariate PROBIT model

Abstract

Objectives

This study aims to develop a prognostic model, based on the video Head Impulse Test (vHIT)-Vestibulo-Ocular Reflex (VOR) gain, of vestibular loss after an acute Unilateral Vestibulopathy (aUVP) predicting the probability of an objective and significant recovery.

Design

Data gathered prospectively in patients with aUVP were re-analyzed. After an exploratory cluster analysis, a multivariate PROBIT regression model was used to objectively pinpoint the value of vHIT-VOR gain at baseline that predicts a significant recovery after 10 weeks. Significant recovery was defined as a vHIT-VOR gain superior or equal to 0.70 after 10 weeks.

Results

A total of 48 subjects were included. The final model identified 2 thresholds, respectively, at 0.53 and 0.74, to classify severe (vHIT-VOR gain ≤0.53), moderate (0.53 < vHIT-VOR gain ≤0.74), and non-significant (0.74 < vHIT-VOR gain) vestibular loss. The probability of a significant recovery was 20% ((Prediction Interval 95% (PI) [0%–43%]) for a gain of 0.50 at baseline, while it doubles (45% (PI [0%–63%]) for an initial gain of 0.60. Our final model demonstrated a good area under the curve (AUC) of 0.85. The corrected AUC after bootstrap resampling was 0.81 (CI 95% [0.74, 0.89]). A good Brier score of the predicted probabilities was also obtained, at 0.15.

Conclusions and relevance

For the first time, this paper proposes a model that objectively defines the severity of vestibular loss at baseline after an aUVP. This research lays the groundwork for future studies to validate our prognostic model and perform more precise analyses of vestibular compensation patterns.

Keywords

vestibular loss acute unilateral vestibulopathy predictive model VOR vestibular neuronitis

Introduction

Acute unilateral vestibulopathy (aUVP) is an abrupt unilateral loss of vestibular function. Its diagnosis is mainly made by exclusion of life-threatening etiologies such as stroke.^1,2 With an approximate incidence rate of 3.5 to 15.5 per 100,000 persons per year,³ it is one of the most common vestibular etiologies that lead to emergency consultations.^4,5

Assessing vestibular function in acute vestibular disorders in a quantitative way is essential to assist in diagnosis and management.^3,6 A readily available and efficient tool to help practitioners in this task is the video head impulse test (vHIT).^7–11 For the diagnosis of aUVP, the Barany society recommends to evaluate the vestibulo-ocular reflex gain with the vHIT (vHIT-VOR gain)³ before other tests. A significant unilateral vestibular loss (VL) is defined in the Barany society criteria as a vHIT-VOR gain of one of the lateral semicircular canals (lat-SCCs) < 0.70; or vHIT-VOR gain discrepancy between the two lat-SCCs > 0.30³. However, assessing vestibular function alone is insufficient for a clear diagnosis of aUVP. It is indeed frequent that, even after a thorough assessment, the diagnosis remains uncertain. For this reason, the Barany Society proposed a sub-entity, the “probable acute vestibulopathy.”³ It is used, for instance, when there is a significant VL but the symptoms do not meet the diagnostic criteria.

On the other hand, the recovery of the vHIT-VOR gain, in the late stage after onset, is used as a marker of the vestibular compensation.^6,10,11 The mechanisms underlying vestibular compensation are yet partly known and include neurogenesis, neuroplasticity, sensory substitution, and habituation.^12–14 Lacour defines a “restoration” of the function when “the lost function is recovered with the original structural elements and operating mode, before the vestibular damage.”¹² Nonetheless, even after restoration of the vHIT-VOR gain, “ad integrum,” individuals with aUVP can experience dizziness at 3 months post-onset.¹⁵

If a total restoration does not appear as a sufficient condition for a complete resolution of vestibular symptoms, the severity of vestibular loss at baseline seems to be relevant to predict the outcome of the compensation. Studying the effect of an early rehabilitation treatment on vestibular function, Lacour and his team found interesting outcomes. Their results revealed that two conditions were necessary for a thorough functional recovery after an aUVP: an early individualized physiotherapy treatment (starting before 15 days post-onset) and a residual function (measured by the vHIT-VOR gain at baseline) of at least 0.20. In this study, severe vHIT-VOR gain loss at baseline was determined by a cluster analysis.¹⁶

Given the remaining uncertainty to diagnose an aUVP and the need to customize rehabilitation,¹⁷ exploring the role of the vestibular loss is of great significance. Characterizing the severity of VL with the vHIT-VOR gain could offer a good option to stratify individuals with aUVP into categories (such as severe, moderate, and non-significant). Several authors proposed this solution to support clinical decision-making, personalize treatment, and further investigate the mechanism underlying vestibular compensation.^18,19

As noted previously, the vHIT-VOR gain is a dynamic outcome measure that changes over time. Following an aUVP, the characterization of VL and its temporal evolution remains a significant unanswered question.

Although only a small number of authors have attempted to address these questions, the reliability of the methods employed to establish the categories, in parallel with the predictions of recovery, can be regarded as debatable (arbitrary categorization, absence of prediction interval).^19,20 Therefore, the present study aims to develop a model to categorize the severity of VL after an aUVP that predicts the probability of an objective and significant recovery of the vHIT-VOR gain.

Materials and methods

This study was written following the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD)-Cluster guidelines²¹ and risk of bias was assessed using the Prediction model Risk Of Bias ASsessment Tool (PROBAST) checklist.²²

Participants and data collection

We performed a secondary analysis on data collected prospectively during the multicenter prospective study of Van Laer et al.¹⁸ In this previous study, patients were included if diagnosed with unilateral vestibular loss including iatrogenic, traumatic, and inflammatory disorders.^18,23 VOR gain data were used from two different time points: up to 4 weeks post-onset and at 10 weeks post-onset. Objective and subjective vestibular functions were evaluated with the vHIT-VOR gain and Dizziness Handicap Inventory (DHI), respectively. Data were collected between June 2021 and September 2023 in three centers. All subjects received a home exercise program that included gaze stabilization, balance, and habituation exercises.¹⁸ The exercises given were tailored, based on the patient’s complaints. The lateral vHIT-VOR gain was measured with the ICS-Impulse vHIT; Otometrics/Natus, Denmark,²⁴ which incorporates a single embedded camera positioned to record the right eye. All vHITs were performed by the same researcher, and impulses with eye blinks were excluded. The camera goggles were tightly secured to minimize slippage. A total of 15 valid impulses were recorded on each side, with head velocity ranging from 120 to 250°s^-1. This study was approved by the ethical committee of the University of Antwerp and the Jessa Hospital (21/12/181), and the study protocol was registered (ID: NCT04979598). All participants provided their written informed consent.

Only individuals with the diagnosis of vestibular aUVP, also called vestibular neuritis, as defined by the Barany Society in 2022,³ were selected for this secondary analysis. Therefore, were included, patients with: (1) an acute or subacute onset of vertigo (spinning or non-spinning), (2) lasting for at least 24 h, (3) presenting a spontaneous nystagmus enhanced by removal of visual fixation, and beating toward the healthy side, (4) with a reduced VOR function measure at the vHIT or caloric testing at the lesion side, and (5) no evidence for acute central neurological, otological or audiological symptoms.

Data preparation, outliers’ management, and sample size

From the initial dataset, the following variables were collated: demographic characteristics (age and gender); side of the lesion, DHI at baseline and 10 weeks post-onset, vHIT-VOR gain assessed in the first 4 weeks post-onset (aVOR1) and at 10 weeks post-onset (aVOR2).

Although the initial dataset also included 6-month follow-up data, only the 10-week follow-up data were used in the present study, for two reasons. Firstly, to ensure data completeness. Indeed, Van Laer et al.²³ reported a higher dropout rate at 6 months than at 10 weeks, and thus, a significant amount of missing data at 6 months. Secondly, compensatory mechanisms have been shown to occur predominantly by 10 weeks post-onset,^23,25 with a plateau effect effectively reached between 9 and 12 weeks after onset.²⁶ Hence, the 10-week data appeared to be the best candidate from a clinical and statistical standpoint.

A clinically significant recovery of the VOR gain was defined in line with the Barany Society recommendation,³ as a gain greater than or equal to 0.70. Consequently, $a V O R 2 \geq 0.70$ was encoded as 1 and $a V O R 2 < 0.70$ was encoded as 0.

Missing values previously identified in the study of Van Laer et al.. were explained by logistical constraints (i.e., distance to travel to the lab or fixed timeslots in which the participant was not available) for patients with aUVP. They were then considered missing at random.²⁷ Since they represented less than 5% of our study sample, a listwise deletion was applied. We considered they would not have significant clinical relevance nor would they have an implication for estimation bias regarding the design and the purpose of our study.²⁸ Consequently, multiple imputation was not applied to handle missing values.

The first author assessed data quality and the risk of bias in the original dataset (see Supplement 1). Potential error outliers²⁹ were then identified by plotting the vHIT-VOR gain at baseline and post-onset in a scatter plot (see Supplement 2). In subjects with very low vHIT-VOR Gain at baseline (<0.2), the DHI scores were expected to be greater than 60.³⁰ Inconsistencies between these two measurements were used as a proxy to identify true error outliers, as they might reflect error in collecting or manipulating data. The detection of outliers was completed by a quantitative technique using the z-score methodology.³¹ The z-scores were calculated for the vHIT-VOR gain at baseline. Outliers were detected when the z-score was below −3 or above 3. The combination of the two techniques allowed to remove outliers from the dataset when respecting the rule of 10 subjects per variable,³² even though there is no consensus on the best way to define the sample size in multivariate non-linear analysis.^21,33

Statistical analyses

A cluster and a PROBIT regression analysis were applied. The cluster analysis was used first to outline the initial cut-off points. Then, a PROBIT analysis was used to refine this cut-off point based on the probability of recovery from the vestibular loss.

All analyses were performed with MATLAB (version R2023b, MathWorks, Natick, MA, USA) with the Statistics and Machine Learning Toolbox, between February and June 2025.

Cluster analysis

We used a k-Means cluster analysis as a first exploratory test to group individuals regarding their initial vestibular loss. We defined the number of subgroups based on previous findings. A review of the literature was conducted to identify subgroups of individuals meeting the following criteria: 1) individuals within a subgroup must have similar long-term recovery outcomes after an aUVP; and/or (2) individuals from two different subgroups must respond differently to treatment (physical, medical, or surgical); and (3) a clear threshold regarding vHIT-VOR gain must be identified between two groups by the authors.

The first group consisted of individuals with “severe loss” with a vHIT-VOR gain below 0.2,¹⁶ 0.4,¹⁹ or 0.5.²⁰ Those individuals have a worse prognosis for recovery of the gain compared to others, i.e. they have less chance to recover the initial loss²⁰ and they are likely to recover incompletely.^16,20 Studies commonly distinguish individuals with severe vestibular loss from those with non-severe loss, whom we refer to as having moderate loss. Individuals in the moderate loss group are generally likely to achieve full recovery following vestibular rehabilitation.¹⁶ In addition, a third group is frequently described in the literature as having a “non-significant loss” typically defined by a lateral vHIT-VOR gain above 0.8 or 0.7.^{3,16,18–20,34,35} Although those individuals are often diagnosed with vestibular neuritis or aUVP, their lateral vHIT-VOR gains remain within normal limits at onset. However, they may still exhibit vestibular dysfunction at lower frequencies or in the vertical canals.^36–38 In such cases, the bithermal caloric test is recommended to detect vestibular loss.³ Despite receiving relatively little attention, recent studies suggest that these individuals may also experience persistent dizziness in the chronic stage.^15,18,39

Thus, the k-Means analysis helped to characterize those three subgroups based on the value of the vHIT-VOR gain at baseline, the aVOR1. Squared Euclidean distance was used as the distance metric to assign data points to clusters in the k-Means algorithm. The random number generator was initialized with the default seed (seed = 0) to allow reproducibility. The maximum number of iterations was set to 100. However, the algorithm stopped earlier if the change in the total within-cluster sum of point-to-centroid distances between two iterations was less than 1e^-4.

The two thresholds, respectively called iT1 and iT2, were defined visually (see Supplement 2).

Regression analysis: PROBIT model

Predictors and outcomes

After the cluster analysis, a PROBIT regression analysis was performed.⁴⁰ PROBIT stands for “probability unit.” It is a predictive model that allows to treat binary dependent variables such as

P [y = 1 | x] = Φ (u)

where

Φ (u)

is the standard normal cumulative distribution function (the link function) which satisfies the inequality 0 ≤ Φ(u) ≤ 1 for any value of

u

. The PROBIT model was first described in 1934⁴¹ and works the same way as LOGIT model, which is commonly referred to as logistic regression.⁴⁰ However, Hahn and Soyer have shown that using a PROBIT link function improves the model fit compared to the logistic function, in case of multivariate models and when the independent variable presents few extreme values, which is our situation (see Supplement 3).⁴²

The purpose of the PROBIT analysis was to predict the probability of significant recovery (aVOR2 > 0.70), given the value of aVOR1, assuming that the response is different if individuals are in a subgroup, whether another. Therefore, we applied a multivariate PROBIT model^43,44 in which the predictor variable _ $a V O R 1$ _ can fall into several categories: severe (GA), moderate (GB), non-significant (GC). Those three categories were used as predictors of our final model.

The dependent binary variable was the significant recovery of aVOR2. Hence, the outcome was the probability of significant recovery of the vHIT-VOR gain, that is, $P (y \geq 0.7)$ with $y = a V O R 2$ .

A first univariate PROBIT analysis was performed, referred to as model A. Therein, the predictor was the $a V O R 1$ with no categories. Then, the multivariate PROBIT analysis, model B, was fitted with the three predictors as described above. Thus, the two final thresholds that define the bounds of the three groups were calculated based on iT1±0.10 and iT2±0.10.

Adjustments of the model, model performance, and internal validation

To refine the threshold values based on those that best fit the data, we ran the model for 100 values of the threshold 1 and the threshold 2, which were included in the intervals [iT1-0.10; iT1+0.10] and [iT2-0.0; iT2+0.10], respectively. We selected the best model in terms of R² adjusted. The corrected Akaike’s Information Criterion (AICc) was also used for model comparison. Furthermore, model discrimination was assessed with receiver operator characteristics and was considered as good if the area under the curve (AUC) was greater than 0.70.⁴⁵ Model evaluation was completed with the computation of the Brier score (interpreted as accurate if < 0.2).³³ Only internal validation was performed with a bootstrap resampling (n = 1000) to evaluate model optimism on the AUC.

Results

Patient characteristics and risk of bias

Of the 48 subjects included in the statistical analysis, 22 (46%) were female and 23 (48%) had right vestibular neuritis (Figure 1). The average age was 55 years (standard deviation (SD) 18 years). The average VOR gain was 0.63 (SD 0.22) at baseline and 0.75 (SD 0.25) at 10 weeks. The median timing of the baseline and 10-week assessments was, respectively, 15.5 days (interquartile range (IQR) 8–31 days) and 74 days (IQR 69–78 days).

Figure 1.

Flow chart of inclusion.

PROBAST assessment showed low risks of bias for population, predictors, and outcome domains. A low risk of concern for the applicability to our study was also found (Supplement 1).

K-means clustering and PROBIT analysis

K-means clustering outlined 3 groups along with two initial thresholds iT1 at 0.44 and iT2 at 0.75 (cf Supplement 2). The PROBIT Model A (cf Figure 2) showed a coefficient of determination (R² adjusted) at 0.30 (p-value <0.001). When adding the 2 thresholds, PROBIT Model B improved substantially with an increased R² adjusted at 0.33 and a good discrepancy between models’ AICcs of 13.22 (cf Supplement 4). The final thresholds T1 and T2 were respectively found at 0.53 and 0.74. Following the final model, unilateral VL can be roughly defined: (1) as severe when the aVOR1 is ≤0.53; (2) as moderate when the aVOR1 is between 0.53 and 0.74; (3) as non-significant when the aVOR1 is >0.74. The predicted probability to recover significantly at 10 weeks in severe, moderate, and non-significant loss at baseline was 20–25%, 45–75% and 96–100%, respectively. With an aVOR1 of 0.5, the probability of a significant recovery is about 20% (Prediction Interval (PI) 0%–43%), whereas it increases up to 45% (PI 0%–63%) with a gain of 0.6. (cf Table 1). Besides, model discrimination did not significantly change from model A to B (Figure 3). Corrected AUC after bootstrap resampling was at 0.81 (IC 95% [0.74, 0.89]). Model B also benefits from a good Brier score of 0.15. Additional information regarding the final clusters is provided in Table 2.

Figure 2.

PROBIT models A and B.

Table 1.

Table of predicted probabilities of a significant recovery given the vHIT-VOR gain at baseline using a multivariate PROBIT model.

vHIT-VOR gain at baseline	Predicted probability of a significant recovery in %	Prediction interval
0.2	25	[0; 67]
0.3	23	[0; 50]
0.4	22	[3; 40]
0.5	20	[0; 43]
0.6	45	[22; 63]
0.7	75	[53; 100]

Figure 3.

Discrimination plot for model A and model B.

Table 2.

Characteristics of the final groups.

	Group with severe loss (n = 13)	Group with moderate loss (n = 20)	Group with non-significant loss (n = 15)
Mean age (SD)	53 ( $\pm$ 19)	59 (±18)	50 (±18)
Mean aVOR1 (SD)	0.35 (±0.08)	0.62 (±0.05)	0.88 (±0.12)
Mean aVOR2 (SD)	0.55 (±0.27)	0.74 (±0.15)	0.94 (±0.19)

SD: standard deviation, n: number of subjects per group, aVOR1: vHIT-VOR gain at baseline, aVOR2: vHIT-VOR at 10 weeks post-onset.

Raw data of the PR score (along with raw vHIT-VOR gains) can be found in Supplement 5 but were not included in this analysis.

Discussion

Our study aimed to describe an objective approach to define vestibular loss after aUVP. However, it was important to establish a clinically relevant definition. Therefore, our study provided a predictive model to objectively characterize the vestibular loss given its prognosis of recovery post-onset. Based on our data, a severe loss is considered when the vHIT-VOR gain at baseline is below 0.53, and the predicted probability to recover significantly is between 20 and 25%. A moderate loss is defined when this gain is between 0.53 and 0.74, along with a predicted probability of recovery between 45 and 75%. Finally, a non-significant loss is defined when the gain is above 0.74, and the predicted probability of recovery is between 96 and 100%.

In a previous study, Lacour et al.,¹⁶ suggested that the threshold for a severe VL was 0.2. The latter was identified by a k-Means cluster analysis using the vHIT-VOR gain at baseline, resulting in two groups. On the contrary, our findings suggest that the cut-off for a severe vestibular loss is circa 0.5. An explanation for such a discrepancy is the difference of measurement tools. Indeed, Lacour et al.¹⁶ used the VHIT Ulmer (SYNAPSYS, Marseille), which is a remote infrared camera whereas Van Laer et al.²³ used an ICS-Impulse vHIT (Otometrics/Natus, Denmark), which uses an embarked camera. Those two devices can show several differences in the VOR outcome^46,47 due mainly to technical differences. Firstly, remote cameras offer the possibility to measure binocular movements to compute vHIT-VOR gain for lat-SCCs as recommended by Weber et al.⁴⁸ Although the ICS-Impulse vHIT has benefited from a validation with a scleral search coil⁴⁹(the gold standard for VOR measurement), it solely records one eye movement during assessment, leading to potential inaccuracy in the measurement in the horizontal plane. Indeed, in normal subjects, the authors observed a significantly larger VOR gain in the ipsilateral direction of the eye recorded.^49,50 Secondly, embarked cameras present a risk of goggle slippage, causing an artificial increase of the VOR gain.⁵¹ Thus, minimizing artefacts using those embarked devices demands from the clinician substantial expertise and training.^52,53 This is particularly true owing to the fact that hand position and low velocity at head impulses can also significantly impact the outcome of the VOR gain in aUVP.⁴⁸ However, our findings remain aligned with the thresholds defined using data collected with embarked cameras and performed by a single experienced operator.^19,20 Therefore, we recommend conducting future studies to replicate this protocol using data from the two types of camera separately.

Also, in our study, as in Lacour’s,¹⁶ the inclusion at baseline may appear enlarged. Indeed, we respectively included data from patients up to 4 weeks after onset and up to 42 days.¹⁶ Subsequently, the included patients are at different stages of compensation of their VOR gain. Due to the constraints of the initial multicenter data collection protocol, it was unfortunately not possible to reanalyze the data using a narrower or more clinically optimal window (such as 3 days post-onset). Despite this heterogeneity, we believe that our findings are close to clinical reality since patients do not always receive a diagnosis at the time of the vertigo insult. Indeed, in a cross-sectional study conducted in Switzerland,⁵ of the 1535 individuals assessed at the emergency department, 148 were diagnosed with aUVP at the acute phase and 35 during a follow-up assessment. Hence, almost 20% of the total number of cases of aUVP were diagnosed retrospectively. This delay in the diagnosis of aUVP is not limited to emergency settings, but also occurs on a broader scale in primary care.⁵⁴ One possible explanation for this phenomenon is that, for some individuals, the vHIT-VOR gain deficit is only detected after a delay.⁵⁵ Another reason could be that it is also often difficult for front-line practitioners to make a diagnosis in the presence of acute vestibular syndrome (AVS).⁵⁶ Some prefer to refer patients to specialists, thereby delaying the initial vHIT assessment by several days or even weeks. Thus, we believe our protocol to be similar to daily routine settings.

In addition to providing an objective definition of the severity of VL, our study provides information about the likelihood of recovery from this loss. To our knowledge, only one other study has quantified the long-term recovery of vestibulo-ocular reflex gain. This retrospective study was conducted by Buki et al.²⁰ and analyzed the vHIT-VOR gain of 44 patients with aUVP graphically. The authors created a scatter plot with baseline gain values on the x-axis, post-onset gain values on the y-axis, and the bisector of the axes. The latter was used as a “predictive model.” The threshold for the severity of the VOR gain loss was defined arbitrarily at 0.5²⁰. No information regarding the model’s goodness of fit, discrimination, or prediction interval was provided. Therefore, our model is the first to integrate predictions of recovery after an aUVP with performance parameters.

In general, combining diagnosis with prognosis establishes a framework of mutual understanding for patients and clinicians.⁵⁷ In this context, the presence of the disease takes on a different meaning for patients, who can now more easily envisage their future with the disease. This change in the approach to pathology, whereby “a useful diagnosis is defined by patient prognosis,”⁵⁷ is also promoted by some authors, as it takes a holistic approach to the patient. Therefore, by defining the severity of the initial VL in relation to the likelihood of significant functional recovery, our model provides patients with aUVP with a better understanding of their disease.

The main strength of this study is our approach to defining the severity of vestibular loss. It is practical and simple for clinicians and researchers to apply. In fact, the parameters of our model (i.e., the vHIT-VOR gain used as predictors and outcomes) are already used in daily routine assessments for the management of patients with dizziness. It is used for diagnostic purposes as well as monitoring.^10,11 Furthermore, the vHIT is a low-cost, quick, and reliable device that can be easily used in emergency settings.^58–60 In AVS, it is even recommended to start the assessment with this tool.⁶¹ Therefore, the methodology we used will be easy to replicate and adapt for further validation studies.

Another strength of this study is the high quality of the initial dataset. Data have been collected recently, prospectively and in multiple centers.¹⁸ This approach reduced multiple risks of bias, notably risks of missing or redundant values, heterogeneity in assessment, in treatment strategies, and in individuals.

Using a multivariate PROBIT model allowed improvement of the AICc and the R² adjusted. Values between 0.2 and 0.4 are often considered a good fit. We believe our model is robust enough regarding the calibration and discrimination outcomes. Nevertheless, the R² adjusted should be carefully interpreted, given the fact that there is no consensus on the best way to calculate this coefficient in cases of multivariate PROBIT regression.^62,63 Overfitting risk related to dichotomization of continuous predictors is also a matter of concern^33,64 and represents a limitation of our study. Especially seeing that the rule of the 10 subjects per variable is debatable. We recommend that further studies anticipate calibration et discrimination parameters beforehand to address these issues.⁶⁵

In our analysis, we focused on the lateral semicircular canal VOR gain as it is the most widely used parameter for assessing vestibular function and is clinically relevant in the context of unilateral loss.^3,6,8,19,35 However, the vestibular system also includes the vertical semicircular canals and the otolith organs, each playing a distinct role in vestibular compensation and responding differently depending on the topography of the aUVP.^37,66,67 We do not have corresponding metrics for these parameters available in this cohort. We recognize this as a limitation and suggest future research to incorporate more of these parameters in an updated version of our model.

A further limitation is that saccade metrics, including frequency, amplitude, latency, and dispersion (as measured by the Perez-Rey (PR) score), were not incorporated into the present predictive model. However, these parameters have demonstrated significant prognostic value in patients with aUVP. Saccade amplitude, frequency, and latency have been shown to be correlated with VOR gain and compensation status.^35,68 Additionally, the PR score, which quantifies the temporal organization of catch-up saccades, has been established as a sensitive marker of vestibular compensation independently of gain values. A validated cutoff of value for distinguishing compensated from uncompensated states was established at 55.⁶⁹ If this metric shows progressive improvement throughout the recovery trajectory, it did not correlate significantly with patient-reported outcomes in the previous study of Van Laer et al.¹⁸ (see Supplement 5). Therefore, we did not include the latter parameter in our multivariate analysis to preserve statistical power given the limited sample size. Future investigations with adequately powered cohorts should integrate saccade metrics into an updated prognostic model, as their inclusion may enhance predictive accuracy for clinical recovery and compensation outcomes following aUVP.

Another limitation of this model is the management of outliers and missing values. PROBIT models assume that the error term follows a normal distribution. Indeed, this assumption leads to a large impact of extreme values on the model.⁷⁰ In this study, error outliers were identified in regard to clinical coherency, using the DHI as a proxy, and other types of outliers were identified with the mean of the z-score. This dual method allowed to retrieve the main outlier while keeping the robustness of the model. This option appeared to be the best in terms of limited sample size (n = 50),⁷¹ given the completeness and quality of the data set. Nevertheless, in subsequent validation studies, it will be of great interest to assess missing values and outliers, and to plan customized management.^28,29

To summarize, our limitations principally stem from the sample size. Evaluation of this model on a larger sample would significantly help to mitigate those limitations and refine the predicted probability of recovery with greater certainty.

Nevertheless, several applications of our model outcome can be described. Thresholds of vHIT-VOR gain can guide clinical decisions in aUVP. Distinguishing severe, moderate, and non-significant vestibular deficits could help prioritize appropriate interventions. For instance, patients with an initial VOR gain below 0.53 might require more intensive, early vestibular physiotherapy. In contrast, patients with an initial VOR gain above 0.74 might benefit more from early behavioral-cognitive therapy that addresses psychological factors and coping mechanisms. This approach will ensure that treatment is matched to the severity of vestibular dysfunction and optimizes rehabilitation outcomes. We believe that our findings can be used in clinical practice to estimate the probability of VOR gain recovery, but they cannot guarantee full subjective recovery, as other factors, such as sensory reweighting or the presence of psychological factors, are known to influence subjective complaints.^{15,18,25,72,73}

Additionally, informing the patient about the severity of their VL has a dual benefit. Firstly, it supports the delivery of individualized and more accurate information regarding their diagnosis. Although our model predictions align with clinical expectations, we believe that the delineation of explicit, data-driven thresholds strengthens prognostic objectivity and clinical utility. In daily practice, qualitative assessments alone often prove insufficient for informing prognostic decisions, as clinicians may find it challenging to convert general impressions (if the gain is low, the recovery might be slower) into specific recommendations. Associating defined VOR gain values with quantifiable recovery probabilities enables a standardized approach to patient counselling and facilitates decision-making regarding the vestibular rehabilitation interventions Significant impacts on the outcome of aUVP can also be expected. Indeed, patients with AVS tend to adopt more positive attitudes toward dizziness when they are diagnosed with an aUVP than when they are diagnosed with a central vestibular disorder.⁷⁴ This phenomenon is referred to as the “therapeutic effect of the diagnosis.” Subsequently, the evolution of dizziness is positively impacted when patients are informed of the peripheral origin of their acute vestibular syndrome. Improved adherence to treatment should also be an outcome of individualized diagnosis.^75–77

Secondly, it facilitates the development of better-tailored treatment plans based on the likelihood of recovery from neuritis. Indeed, recent findings on modular treatment strategies, whether through physiotherapy, cognitive therapy, or both, revealed that customized interventions reduced the impact of dizziness in the short term.^78,79 Finally, since psychological traits have been shown to be high risk factors to chronic dizziness, even when the VL is minimal,^{18,73,80–83} we anticipate that improving tailored diagnosis and prognosis will not only impact outcomes in short term but might also reduce risk of chronification of the symptoms.

Conclusion

In this paper, we propose for the first time a predictive model to objectively define the severity of vestibular loss after an aUVP. In addition, we provide the probability of recovery of the vHIT-VOR gain after 10 weeks, given the vHIT-VOR gain at baseline. If our final multivariate PROBIT model shows improved performance over the initial model, external validation is still required to increase the adequacy of our predictions. Further study should be undertaken with a large sample size. However, this study proposes an objective classification of individuals with aUVP that remains understandable to the patient. A vestibular loss is considered: (1) as severe when the vHIT-VOR gain is ≤0.53; (2) as moderate when the vHIT-VOR gain is between 0.53 and 0.74; and (3) as non-significant when the vHIT-VOR gain is >0.74. The predicted probability of recovering significantly at 10 weeks in severe, moderate and non-significant loss at baseline were ranging 20–25%, 45–75% and 96–100%, respectively. This study lays the foundation for future, more precise analyses of vestibular compensation patterns.

Supplemental material

Supplemental material - Using the probability of recovery from an acute unilateral vestibulopathy to determine the severity of the vestibular loss: Development of a multivariate PROBIT model

Supplemental material for Using the probability of recovery from an acute unilateral vestibulopathy to determine the severity of the vestibular loss: Development of a multivariate PROBIT model by Gabrielle Vassard-Yu, Lien Van Laer, Ann Hallemans, Luc Vereeck, and Vincent Van Rompaey in Journal of Vestibular Research

Supplemental material

Supplemental material - Using the probability of recovery from an acute unilateral vestibulopathy to determine the severity of the vestibular loss: Development of a multivariate PROBIT model

Supplemental material

Supplemental material - Using the probability of recovery from an acute unilateral vestibulopathy to determine the severity of the vestibular loss: Development of a multivariate PROBIT model

Supplemental material

Supplemental material - Using the probability of recovery from an acute unilateral vestibulopathy to determine the severity of the vestibular loss: Development of a multivariate PROBIT model

Supplemental material

Supplemental material - Using the probability of recovery from an acute unilateral vestibulopathy to determine the severity of the vestibular loss: Development of a multivariate PROBIT model

Footnotes

Acknowledgments

The authors wish to thank Massimiliano Moda from the Department of Mathematics of UAntwerp for his help in writing the MATLAB code. The STATUA team for its help while revising the manuscript; all the participants for their time and willingness to participate; and all referring doctors from the Antwerp University Hospital, Gelre Hospital, the Jessa Hospital (Hasselt, Belgium), the Rehabilitation Center Sint-Lievenspoort (Ghent, Belgium), and the AZ Turnhout Hospital (Turnhout, Belgium) for their help in the recruitment of participants and their guidance throughout our research project. VVR is an FWO Senior Clinical Investigator (18E2524N).

ORCID iD

Gabrielle Vassard-Yu

Ethical considerations

The study was conducted in accordance with the Declaration of Helsinki. Ethical review and approval were waived for this study due to the retrospective nature of the analysis.

Author contributions

Conceptualization, G.V-Y.; methodology, G.V-Y.; software, G.V-Y and Massimiliano Moda; formal analysis, G.V-Y; resources, G.V-Y and L.V-L; writing—original draft preparation, G.V-Y.; and writing—review and editing, G.V-Y., L.V-L., V.V.R., L.V., and A.H. All authors have read and agreed to the published version of the manuscript.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was funded by the EUROPEAN UNION, HORIZON EUROPE Marie Sklodowska-Curie Actions, grant number 101120139.

Declaration of conflicting interests

The authors declare no conflicts of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

Data Availability Statement

The dataset used for this study is referred to in Supplement 5.*

Use of AI

No generative AI tools were used in the preparation of this manuscript for content generation, data analysis, or figure creation. We want to acknowledge that we used DeepL Write, an AI tool, to correct the grammar in parts of this manuscript.

Supplemental material

Supplemental material is available online.

References

Kattah

Talkad

Wang

, et al. HINTS to diagnose stroke in the acute vestibular syndrome: three-step bedside oculomotor examination more sensitive than early MRI diffusion-weighted imaging. Stroke 2009; 40: 3504–3510.

Tarnutzer

Berkowitz

Robinson

, et al. Does my dizzy patient have a stroke? A systematic review of bedside diagnosis in acute vestibular syndrome. CMAJ (Can Med Assoc J) 2011; 183: E571–E592.

Strupp

Bisdorff

Furman

, et al. Acute unilateral vestibulopathy/vestibular neuritis: diagnostic criteria. J Vestib Res 2022; 32: 389–406.

Zwergal

Dieterich

. Vertigo and dizziness in the emergency room. Curr Opin Neurol 2020; 33: 117–125.

Comolli

Korda

Zamaro

, et al. Vestibular syndromes, diagnosis and diagnostic errors in patients with dizziness presenting to the emergency department: a cross-sectional study. BMJ Open 2023; 13: e064057.

Swamy Suman

Kumar Rajasekaran

Yuvaraj

, et al. Measure of central vestibular compensation: a review. J Int Adv Otol 2022; 18: 441–446.

Halmagyi

Chen

MacDougall

, et al. The video head impulse test. Front Neurol 2017; 8: 258.

Strupp

Grimberg

Teufel

, et al. Worldwide survey on laboratory testing of vestibular function. Neurol Clin Pract 2020; 10: 379–387.

MacDougall

Weber

McGarvie

, et al. The video head impulse test: diagnostic accuracy in peripheral vestibulopathy. Neurology 2009; 73: 1134–1141.

10.

Manzari

Burgess

MacDougall

, et al. Vestibular function after vestibular neuritis. Int J Audiol 2013; 52: 713–718.

11.

Manzari

Burgess

MacDougall

, et al. Objective verification of full recovery of dynamic vestibular function after superior vestibular neuritis. Laryngoscope 2011; 121: 2496–2500.

12.

Lacour

Bernard-Demanze

. Interaction between vestibular compensation mechanisms and vestibular rehabilitation therapy: 10 recommendations for optimal functional recovery. Front Neurol 2015; 5: 285. https://doi.org/10.3389/fneur.2014.00285 (accessed 11 December 2023).

13.

Lacour

Helmchen

Vidal

P-P

. Vestibular compensation: the neuro-otologist’s best friend. J Neurol 2016; 263: 54–64.

14.

Tighilet

Bordiga

Cassel

, et al. Peripheral vestibular plasticity vs central compensation: evidence and questions. J Neurol 2019; 266: 27–32.

15.

Armato

Dumas

Perottino

, et al. Determination of recovery by total restitution or compensation using multifrequency vestibular tests and subjective functional scales in a human model of vestibular neuritis. Audiol Res 2024; 14: 958–982.

16.

Lacour

Thiry

Tardivet

. Two conditions to fully recover dynamic canal function in unilateral peripheral vestibular hypofunction patients. J Vestib Res 2021; 31: 407–421.

17.

Hall

Herdman

Whitney

, et al. Vestibular rehabilitation for peripheral vestibular hypofunction: an updated clinical practice guideline from the academy of neurologic physical therapy of the American physical therapy association. J Neurol Phys Ther 2022; 46: 118–177.

18.

Van Laer

Hallemans

, et al. Predictors of chronic dizziness in acute unilateral vestibulopathy: a longitudinal prospective cohort study. Otolaryngol--Head Neck Surg Off J Am Acad Otolaryngol-Head Neck Surg 2024. https://doi.org/10.1002/ohn.964, Epub ahead of print 3.

19.

Alfarghal

Algarni

Sinha

, et al. VOR gain of lateral semicircular canal using video head impulse test in acute unilateral vestibular hypofunction: a systematic review. Front Neurol 2022; 13: 948462.

20.

Büki

Hanschek

Jünger

. Vestibular neuritis: involvement and long-term recovery of individual semicircular canals. Auris Nasus Larynx 2017; 44: 288–293.

21.

Debray

TPA

Collins

Riley

, et al. Transparent reporting of multivariable prediction models developed or validated using clustered data: TRIPOD-cluster checklist. Br Med J 2023; 380: e071018.

22.

Wolff

Moons

KGM

Riley

, et al. PROBAST: a tool to assess the risk of bias and applicability of prediction model studies. Ann Intern Med 2019; 170: 51–58.

23.

Van Laer

Hallemans

Janssens de Varebeke

, et al. Compensatory strategies after an acute unilateral vestibulopathy: a prospective observational study. Eur Arch Otorhinolaryngol 2024; 281: 743–755.

24.

Bansal

Sinha

. Assessment of VOR gain function and its test-retest reliability in normal hearing individuals. Eur Arch Otorhinolaryngol 2016; 273: 3167–3173.

25.

Cousins

Kaski

Cutfield

, et al. Predictors of clinical recovery from vestibular neuritis: a prospective study. Ann Clin Transl Neurol 2017; 4: 340–346.

26.

Vereeck

Wuyts

Truijen

, et al. The effect of early customized vestibular rehabilitation on balance after acoustic neuroma resection. Clin Rehabil 2008; 22: 698–713.

27.

Kang

. The prevention and handling of the missing data. Korean J Anesthesiol 2013; 64: 402–406.

28.

Sperrin

Martin

Sisk

, et al. Missing data should be handled differently for prediction than for description or causal explanation. J Clin Epidemiol 2020; 125: 183–187.

29.

Aguinis

Gottfredson

Joo

. Best-Practice Recommendations for Defining, Identifying, and Handling Outliers. 2013. Accessed March 27, 2025. https://doi.org/10.1177/1094428112470848?casa_token/EEj-uhe6QWgAAAAA/3A9HZSLAY-58dUOFqoZ7jKgF2zUJ3iCJkjTplQRyJ9vknM7cwYJSurpCjuz2kKjzYPlYLGQlMYQE0B (accessed 27 March 2025).

30.

Esteban Sanchez

Martín-Sanz

. Long-term evolution of vestibular compensation, postural control, and perceived disability in a population of patients with vestibular neuritis. J Clin Med 2022; 11: 3941.

31.

Anusha

Anuradha

Murty

PSRC

, et al. Detecting outliers in high dimensional data sets using Z-score methodology. Int J Innovative Technol Explor Eng 2019; 9: 48–53.

32.

Streiner

Kottner

. Recommendations for reporting the results of studies of instrument and scale development and testing. J Adv Nurs 2014; 70: 1970–1979.

33.

Clinical prediction models: a practical approach to development, validation, and updating | SpringerLink. https://doi.org/10.1007/978-3-030-16399-0 (accessed 31 March 2025).

34.

Taylor

Kong

Flanagan

, et al. Prevalence of vestibular dysfunction in patients with vestibular schwannoma using video head-impulses and vestibular-evoked potentials. J Neurol 2015; 262: 1228–1237.

35.

Psillas

Petrou

Printza

, et al. Video head impulse test (vHIT): value of gain and refixation saccades in unilateral vestibular neuritis. J Clin Med 2022; 11: 3467.

36.

Bartolomeo

Biboulet

Pierre

, et al. Value of the video head impulse test in assessing vestibular deficits following vestibular neuritis. Eur Arch Otorhinolaryngol 2014; 271: 681–688.

37.

Taylor

McGarvie

Reid

, et al. Vestibular neuritis affects both superior and inferior vestibular nerves. Neurology 2016; 87: 1704–1712.

38.

Albernaz

PLM

Cusin

Ramos

, et al. Vestibular disorders limited to the vertical semicircular canals. Int Arch Otorhinolaryngol 2024; 28: e587–e591.

39.

Kim

D-Y

Hwang

, et al. Vestibular neuritis with minimal canal paresis: characteristics and clinical implication. Clin Exp Otorhinolaryngol 2016; 10: 148–152.

40.

Hansen Bruce

. Econometrics/Bruce E. Hansen. Princeton etc. Princeton university press, 2022.

41.

Bliss

. The method of probits. Science 1934; 79: 38–39.

42.

Hahn

Soyer

(eds). Probit and logit models: differences in the multivariate realm. Stat Soc; Ser B; 67: 1–12.

43.

Lesaffre

Molenberghs

. Multivariate probit analysis: a neglected procedure in medical statistics. Stat Med 1991; 10: 1391–1403.

44.

Hidalgo

Goodman

. Multivariate or multivariable regression? Am J Publ Health 2013; 103: 39–40.

45.

Hond

AAH

Steyerberg

Calster

. Interpreting area under the receiver operating characteristic curve. Lancet Digit Health 2022; 4: e853–e855.

46.

Felipe

. Comparison of two systems for the video head impulse test (vHIT) for the lateral semicircular canal: description of results from normal and pathological subjects. Arq Neuropsiquiatr 2021; 79: 571–578.

47.

van Dooren

Starkov

Lucieer

FMP

, et al. Comparison of three video head impulse test systems for the diagnosis of bilateral vestibulopathy. J Neurol 2020; 267: 256–264.

48.

Weber

Todd

, et al. Inter-ocular differences of the horizontal vestibulo-ocular reflex during impulsive testing. Prog Brain Res 2008; 171: 195–198.

49.

Effects of device on video head impulse test (vHIT) gain: ingenta connect. https://www.ingentaconnect.com/content/aaa/jaaa/2017/00000028/00000009/art00002 (accessed 13 February 2025).

50.

Striteska

Chovanec

Steinmetzer

, et al. Binocular video head impulse test: normative data study. Front Neurol 2023; 14: 1153102. https://doi.org/10.3389/fneur.2023.1153102, Epub ahead of print, (3 May 2023).

51.

Suh

M-W

Park

Kang

, et al. Effect of goggle slippage on the video head impulse test outcome and its mechanisms. Otol Neurotol 2017; 38: 102–109.

52.

Heuberger

Grill

Saǧlam

, et al. Usability of the video head impulse test: lessons from the population-based prospective KORA study. Front Neurol 2017; 9: 00659. https://doi.org/10.3389/fneur.2018.00659, Epub ahead of print.

53.

Korda

Sauter

Caversaccio

, et al. Quantifying a learning curve for video head impulse test: pitfalls and pearls. Front Neurol 2021; 11. https://doi.org/10.3389/fneur.2020.615651, Epub ahead of print.

54.

Bösner

Schwarm

Grevenrath

, et al. Prevalence, aetiologies and prognosis of the symptom dizziness in primary care - a systematic review. BMC Fam Pract 2018; 19: 33.

55.

Skacik

Sivak

Kurca

. Negative video head impulse test in acute vestibular syndrome does not exclude vestibular neuritis: insights and challenges in diagnosis. Clin Case Rep 2025; 13: e70300.

56.

Zwergal

Mantokoudis

Heg

, et al. What is the current status of primary care in the diagnosis and treatment of patients with vertigo and dizziness in Switzerland? A national survey. Front Neurol 2023; 14: 1254080.

57.

Croft

Altman

Deeks

, et al. The science of clinical practice: disease diagnosis or patient prognosis? Evidence about ‘what is likely to happen’ should shape clinical practice. BMC Med 2015; 13: 20.

58.

Roberts

McGuigan

Infeld

, et al. A video-oculographic study of acute vestibular syndromes. Acta Neurol Scand 2016; 134: 258–264.

59.

Mantokoudis

Tehrani

ASS

Wozniak

, et al. VOR gain by head impulse video-oculography differentiates acute vestibular neuritis from stroke. Otol Neurotol 2015; 36: 457–465.

60.

Chen

Todd

Halmagyi

, et al. Head impulse gain and saccade analysis in pontine-cerebellar stroke and vestibular neuritis. Neurology 2014; 83: 1513–1522.

61.

van Esch

Nobel-Hoff

GEa. J

van Benthem

PPG

, et al. Determining vestibular hypofunction: start with the video-head impulse test. Eur Arch Otorhinolaryngol 2016; 273: 3733–3739.

62.

Hagle

Mitchell

. Goodness-of-Fit measures for probit and logit. Am J Polit Sci 1992; 36: 762–784.

63.

Liu

Zhu

Greenwell

, et al. A new goodness‐of‐fit measure for probit models: surrogate R 2. Br J Math Stat Psychol 2023; 76: 192–210.

64.

Steyerberg

Uno

Ioannidis

JPA

, et al. Poor performance of clinical prediction models: the harm of commonly applied methods. J Clin Epidemiol 2018; 98: 133–143.

65.

Snell

KIE

Archer

Ensor

, et al. External validation of clinical prediction models: simulation-based sample size calculations were more reliable than rules-of-thumb. J Clin Epidemiol 2021; 135: 79–89.

66.

Navari

Casani

. Lesion patterns and possible implications for recovery in acute unilateral vestibulopathy. Otol Neurotol 2020; 41: e250–e255.

67.

Piatti

Conti

Paolocci

, et al. Vestibular evoked myogenic potentials to diagnose vestibular neuritis: a scoping review. Laryngoscope 2025. https://doi.org/10.1002/lary.70214, Epub ahead of print 22.

68.

Aubele

Grill

Eggert

, et al. Symptoms in unilateral vestibular hypofunction are associated with number of catch-up saccades and retinal error: results from the population-based KORA FF4 study. Front Neurol 2023; 14: 1292312. https://doi.org/10.3389/fneur.2023.1292312

69.

Guajardo-Vergara

Perez-Fernandez

. A new and faster method to assess vestibular compensation: a cross-sectional study. Laryngoscope 2020; 130: E911–E917.

70.

Hausman

Abrevaya

Scott-Morton

. Misclassification of the dependent variable in a discrete-response setting. J Econom 1998; 87: 239–269.

71.

Green

SB.

How many subjects does it take to Do A regression analysis. Multivariate Behav Res. 1991; 499–510. https://doi.org/10.1207/s15327906mbr2603_7

72.

Yip

Strupp

. The dizziness handicap inventory does not correlate with vestibular function tests: a prospective study. J Neurol 2018; 265: 1210–1218.

73.

Tschan

Best

Beutel

, et al. Patients’ psychological well-being and resilient coping protect from secondary somatoform vertigo and dizziness (SVD) 1 year after vestibular disease. J Neurol 2011; 258: 104–112.

74.

Prell

Finn

Zipprich

, et al.

What predicts improvement of dizziness after multimodal and interdisciplinary day care treatment?

J Clin Med 2022; 11: 2005.

75.

Lee

Jeon

E-J

. Adherence to customized vestibular rehabilitation therapy: influencing factors and clinical implications in vestibulopathy. Front Neurol 2025; 16: 1538989.

76.

Tanaka

Fushiki

Tsunoda

, et al. Effect of vestibular rehabilitation program using a booklet in patients with chronic peripheral vestibular hypofunction: a randomized controlled trial. Prog Rehabil Med 2023; 8: 20230002.

77.

Lee

Smi

C-K

. Factors influencing adherence to vestibular rehabilitation exercise program in patients with dizziness. Korean J Adult Nurs 2014; 26: 434–443.

78.

Liu

, et al. Optimizing vestibular neuritis management with modular strategies. Front Neurol 2023; 14: 1243034. https://doi.org/10.3389/fneur.2023.1243034

79.

Ding

C-R

Gao

Y-Q

Zhou

Y-J

, et al. Advantages of short-term personalized vestibular rehabilitation at home guided by professional therapist for treatment of decompensated vestibular vertigo. Curr Med Sci 2021; 41: 687–694.

80.

Feng

Zang

. The effect of accompanying anxiety and depression on patients with different vestibular syndromes. Front Aging Neurosci 2023; 15: 1208392.

81.

Bronstein

Dieterich

. Long-term clinical outcome in vestibular neuritis. Curr Opin Neurol 2019; 32: 174–180.

82.

Dietzek

Finn

Karvouniari

, et al. In older patients treated for dizziness and vertigo in multimodal rehabilitation somatic deficits prevail while anxiety plays a minor role compared to young and middle aged patients. Front Aging Neurosci 2018; 10: 345.

83.

Adamec

Krbot Skorić

Ozretić

, et al. Predictors of development of chronic vestibular insufficiency after vestibular neuritis. J Neurol Sci 2014; 347: 224–228.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.03 MB

0.04 MB

0.06 MB

0.01 MB

0.00 MB