Prediction Model for Etiologic Differentiation of Isolated Vestibular Syndrome in Emergency Settings

ABSTRACT

Objective

This study aimed to develop and validate a predictive model for differentiating central from peripheral etiologies in patients with isolated vestibular syndrome (VS).

Methods

In this multicenter retrospective cohort study, 506 patients with isolated VS from five hospitals were divided into derivation (n = 301) and validation (n = 205) cohorts. Multivariable logistic regression was performed to determine independent predictors of central VS. These predictors were assigned weights to construct the SAV₃E score. The performance of the SAV₃E was assessed using the area under the curve (AUC), calibration, and decision curve analysis (DCA) and compared with that of the TriAGe+ and STANDING models.

Results

The SAV₃E score incorporated five predictors: absence of vestibular vagal symptoms (OR, 0.233; 95% CI, 0.084–0.648; p = 0.005), prior stroke (OR, 15.204; 95% CI, 4.455–51.884; p < 0.001), ABCD₂ score of 4–7 (OR, 1.903; 95% CI, 1.206–3.004; p = 0.006), central video oculography nystagmus (OR, 38.377; 95% CI, 8.631–170.644; p < 0.001), and positive video head impulse test (OR, 0.078; 95% CI, 0.033–0.188; p < 0.001). It displayed good discriminative performance with AUCs 0.910 and 0.886 in the derivation and validation cohorts, respectively. It outperformed TriAGe+ (AUC: 0.706) and STANDING (AUC: 0.779) models. Furthermore, calibration analysis revealed good model fit across cohorts (Hosmer‐Lemeshow test results: derivation cohort, p = 0.899; validation cohort, p = 0.789). DCA confirmed good clinical utility across a wide range of probability thresholds (derivation cohort: 0.01–0.86, validation cohort: 0.01–1.00).

Conclusion

The SAV₃E score is a validated tool aimed at differentiating central versus peripheral VS, with the potential to improve diagnostic accuracy for urgent etiologies such as stroke.

1. Introduction

Isolated vestibular syndrome (VS)—dizziness, vertigo, and postural instability—is a common presenting symptom in emergency departments. It presents a significant challenge for emergency physicians to differentiate between central (mainly potentially life‐threatening posterior stroke) and peripheral (generally benign) etiologies [1, 2]. Brain magnetic resonance imaging (MRI) is considered the gold standard for diagnosing ischemic stroke. However, its sensitivity for detecting small lesions within 48 h after symptom onset remains suboptimal, and its availability in emergency settings is limited [3]. Non‐contrast cranial computed tomography (CT) remains the first‐line screening tool in emergency settings due to its availability and rapid acquisition time. However, its sensitivity for detecting posterior circulation ischemic strokes is critically limited (16%–30%) [4, 5]. Due to the insufficient diagnostic efficacy of imaging, bedside eye movement examination has become a key clinical method for the early identification of the etiology of VS.

The bedside ocular motor assessment system, represented by the “HINTS” examination, theoretically enables rapid differentiation between central and peripheral etiologies through the triad of head impulse test (HIT), nystagmus, and skew deviation (SD), with a sensitivity of 100% and specificity of 96%, significantly outperforming MRI examinations conducted within 48 h of symptom onset [6]. However, real‐world data reveal a substantial clinical implementation gap, with studies reporting actual HINTS implementation rates of only 5%–19.5% in emergency settings [7, 8, 9]. Moreover, the diagnostic accuracy of HINTS significantly depends on the clinical examiner's expertise, which partly explains the low implementation rates. Clinicians face three core challenges in applying the “HINTS”: insufficient operational confidence, controversies regarding technical efficacy, and deviations from standardized protocol [10, 11, 12, 13, 14]. Therefore, current emergency practice confronts a dual dilemma of low utilization rate and high misjudgment rate in bedside ocular motor assessments. Misinterpretation of ocular movement analysis may lead to clinical decision‐making errors, potentially increasing the risk of delayed diagnosis and treatment of central VS, especially stroke.

Recent years, video eye movement technologies, such as video‐oculography (VOG) and video HIT (vHIT), have been widely adopted in clinical practice. These technologies significantly enhance sign recognition accuracy through quantitative analysis of eye and head movements [15]. Studies indicate that individual or combined VOG/vHIT assessments demonstrate superior capabilities for identifying abnormal eye movements compared to bedside examinations [16, 17, 18, 19]. However, this approach still requires professionally trained personnel for implementation, operation, and interpretation, limiting its accessibility. Developing portable eye‐tracking technology will be vital in the future.

Current risk stratification tools for VS, including the TriAGe+ [15] and STANDING [20] scores, exhibit moderate sensitivity and specificity for central pathologies, emphasizing the need for improved prediction models. Integrating clinical indicators, such as vascular medical history and ABCD₂ scores, with advanced oculomotor parameters may address this gap. However, no validated model systematically incorporates these multimodal parameters. Additionally, population‐specific validation remains insufficient, particularly in Asian cohorts with different prevalences of posterior circulation stroke, which limits the generalizability of existing models.

This study aims to develop and validate a multimodal prediction model (the SAV₃E score) integrating VOG/vHIT parameters and clinical indicators to differentiate the central from the peripheral VS.

2. Methods

2.1. Study Design and Population

We conducted a retrospective multicenter cohort study analyzing consecutive patients with isolated VS (including dizziness, vertigo, and postural instability) from the neurology department across five hospitals in Zhejiang Province (including Zhejiang Provincial People's Hospital, Sir Run Run Shaw Hospital, Affiliated Hospital of Jiaxing University, Tiantai County People's Hospital, and The Second People's Hospital of Yuhang District), China between October 2020 and December 2024. The inclusion criteria were as follows: (1) patients aged 18 years or older; (2) VS lasting more than 5 min and illness duration not exceeding 1 week; (3) pre‐hospital or emergency cranial CT to rule out intracranial hemorrhage or space‐occupying lesions; (4) brain MRI completed within 2 weeks of admission; and (5) completion of VOG confirming spontaneous nystagmus (SN) or gaze‐evoked nystagmus (GEN) within 2 weeks of admission. The exclusion criteria comprised: (1) new focal neurological signs (except ocular motor abnormalities or truncal ataxia); (2) incomplete horizontal vestibulo‐ocular reflex (VOR) data from vHIT; and (3) an undetermined diagnosis for discharge etiology classification. The study protocol was approved by the Ethics Committee of Zhejiang Provincial People's Hospital (No. Zhejiang Provincial People's Hospital Ethics Review 2025 Study No. 037), which waived the requirement for informed consent because of the retrospective nature of the study.

2.2. Data Collection

We systematically collected data on several key areas: (1) Demographics: age and sex; (2) symptomatology: (a) vestibular characteristics, including symptom type (vertigo or non‐vertigo), duration of the last onset, and recurrence pattern (with or without previous attacks); (b) associated manifestations such as vagal (containing nausea, vomiting, and diaphoresis), intensity of vagal (mild, without vomiting; severe, with vomiting), cochlear, and headache symptoms; and (c) syndrome classification into acute VS (AVS) or episodic VS (EVS); (3) Medical history: hypertension, diabetes mellitus, atrial fibrillation, and stroke history; (4) Physical signs: (a) oculomotor findings, including spontaneous nystagmus pattern (peripheral or central) and horizontal HIT results; (b) neurologic assessment with truncal ataxia severity grading (normal/mild/moderate/severe); (c) auditory evaluation for hearing loss; (d) stroke risk stratification using ABCD₂ score; and (5) VOG features: including VOG nystagmus patterns (peripheral or central), vHIT gain asymmetry rate index, and vHIT results.

2.3. Definition of Isolated Dizziness Etiology Classification

The causes of isolated VS were classified into two categories, central and peripheral, based on discharge diagnoses.

Central causes included acute cerebral infarction, posterior fossa tumors, and cranio‐cervical junction anomalies. The latter refers to a spectrum of bone, soft tissue, and neural structural abnormalities, such as basilar invagination, platybasia, atlantoaxial dislocation, tonsillar herniation, and occipito‐cervical fusion [15]. These were identified by brain or cervical MRI. In addition, central causes also included definite vestibular migraine (VM) and encephalitis. VM was diagnosed according to the Barany Society criteria [21]. Encephalitis refers to the inflammation of the brain, which can be ascertained clinically through cerebrospinal fluid parameters, cranial imaging, or electroencephalogram [22].

Peripheral causes included: acute unilateral vestibulopathy (AUVP), diagnosed according to the criteria for acute unilateral vestibulopathy/vestibular neuritis established by the Barany Society [15]; definite Meniere's disease (MD), diagnosed based on the criteria for Meniere's disease established by the Barany Society [23]; sudden sensorineural hearing loss with vertigo (SHL), diagnosed following the 2019 clinical practice guideline for sudden hearing loss [24]; and Ramsay‐Hunt syndrome (RHS), diagnosed according to the European consensus‐based (S₂k) guideline on the management of Herpes Zoster [25].

2.4. Definition of VOG Features

The nystagmus examination results were categorized as either “peripheral nystagmus pattern” or “central nystagmus pattern.” This classification follows the definitions criteria provided by the Barany Society for nystagmus and nystagmus‐like eye movements [26]. The peripheral nystagmus pattern includes horizontal SN or direction‐fixed GEN. In contrast, the central nystagmus pattern consists of vertical or purely torsional SN, as well as direction‐changing GEN.

The results of vHIT were categorized as “positive” or “negative” based on the Chinese expert consensus criteria [27]. A positive result is defined as a horizontal semicircular canal VOR gain of < 0.8 with ipsilateral compensatory saccades. A negative result is defined as a horizontal semicircular canal VOR gain of 0.8 or higher. The gain difference was calculated by taking the absolute value of the differences in bilateral VOR gains. The gain asymmetry ratio was determined using the formula: absolute gain difference divided by the lower gain value.

All video‐oculographic examinations were verified by vestibular function specialists from the participating hospitals in collaboration with Shi Tianming, the Neurology and Vestibular Disorders Specialist for this study.

2.5. Statistics

Categorical variables are reported as numbers with corresponding percentages and analyzed using either the chi‐square test or Fisher's exact test. Continuous variables are presented as mean ± standard deviation (SD) or median (interquartile range, IQR) and analyzed using the independent t‐test or Mann–Whitney U‐test as appropriate.

Data from patients at the Zhejiang Provincial People's Hospital was used for the training cohort, whereas data from patients at four other hospitals was used for the validation cohort. We analyzed the differences between the training cohort and the test cohort.

To establish the predictive model, we compared patients with central and peripheral VS in the training cohort. Next, we used a multivariable logistic regression analysis to determine independent predictors of central VS by inputting variables with a p‐value < 0.05 from the univariate analyses. The odds ratio (OR) and 95% confidence interval (CI) of each significant risk factor were calculated. We assigned integer point values to each independent predictor based on their corresponding β‐coefficients to construct the SAV₃E score. Performances of the model were assessed using the indexes as follows: the area under the receiver operating characteristic (ROC) curve (AUC) for discrimination, the Hosmer‐Lemeshow goodness‐of‐fit test and calibration plot for the calibration, and decision curve analysis (DCA) for clinical usefulness. Besides, the performances of the SAV₃E score were also compared with the STANDING and TriAge+ in the training cohort.

All data were analyzed using SPSS 22.0 and R software packages (version 4.1.2). p value < 0.05 was considered significant.

3. Results

Between October 2020 and December 2024, a total of 506 patients who fulfilled the study criteria were included, with 301 patients in the derivation cohort and 205 patients in the validation cohort, respectively (Figure 1). The characteristics of patients in the derivation and validation cohorts are presented in Table 1.

FIGURE 1.

Flow chart of the patients enrolled in the study. GEN, gaze‐evoked nystagmus; HIT, head impulse test; MRI, magnetic resonance image; SN, spontaneous nystagmus; VOG, video oculography.

TABLE 1.

Comparison of the characteristics between the derivation and the validation cohort.

	Derivation cohort (n = 301)	Validation cohort (n = 205)	p
Demographic characteristics
Gender (female)	114 (37.9%)	76 (37.1%)	0.855
Age	56.41 ± 15.02	54.28 ± 14.39	0.112
Symptomatic features
Vestibular symptom (vertigo)	215 (71.4%)	154 (75.1%)	0.359
Vestibular vagal symptom (absence)	259 (86.0%)	177 (86.3%)	0.925
≤ 24 h	156 (60.2%)	96 (54.2%)
Mild	24 (15.4%)	8 (8.3%)	0.103
Severe	132 (84.6%)	88 (91.7%)
> 24 h	103 (39.8%)	81 (45.8%)
Mild	20 (19.4%)	10 (12.3%)	0.197
Severe	83 (80.6%)	71 (87.7%)
Cochlear symptom	68 (22.6%)	33 (16.1%)	0.073
Headache symptom	15 (5.0%)	5 (2.4%)	0.149
Duration of last onset	24 (5–72)	24 (10–72)	0.013*
≤ 24 h	174 (57.8%)	104 (50.7%)	0.116
> 24 h	127 (42.2%)	101 (49.3%)
Previous attacks	81 (26.9%)	26 (12.7%)	< 0.001*
Vestibular syndrome (acute vestibular syndrome)	277 (92.0%)	203 (99.0%)	< 0.001*
Medical history
Hypertension	131 (43.5%)	79 (38.5%)	0.264
Diabetes	51 (16.9%)	27 (13.2%)	0.249
Atrial fibrillation	5 (1.7%)	3 (1.5%)	0.861
Stroke	20 (6.6%)	14 (6.8%)	0.935
Physical signs
Beside nystagmus pattern
Null	71 (23.6%)	42 (20.5%)	0.002*
Peripheral	211 (56.6%)	162 (79.0%)
Central	19 (6.3%)	1 (0.5%)
Beside HIT
Null	185 (61.5%)	125 (61.0%)	0.743
Negative	17 (5.6%)	15 (7.3%)
Positive	99 (32.9%)	65 (31.7%)
Truncal ataxia
Normal	84 (27.9%)	10 (4.9%)	< 0.001*
Mild	36 (12.0%)	100 (48.8%)
Moderate	54 (17.9%)	46 (22.4%)
Severe	127 (42.2%)	49 (23.9%)
Hearing loss	21 (7.0%)	16 (7.8%)	0.725
ABCD2 score	3 (2–4)	3 (2–3)	< 0.001*
VOG features
VOG nystagmus pattern
Peripheral	282 (93.7%)	198 (96.6%)	0.147
Central	19 (6.3%)	7 (26.9%)
vHIT gain asymmetry rate	0.41 (0.24–0.53)	0.94 (0.40–1.74)	< 0.001*
vHIT a
Negative	64 (21.3%)	36 (17.6%)	0.305
Positive	237 (78.7%)	169 (82.4%)

Abbreviations: HIT, head impulse test; VOG, video oculography.

^a 2 missing.

^* p < 0.05.

3.1. Univariate Analyses of Patients With Central and Peripheral VS in the Training Cohort

Characteristics of patients in the peripheral and central groups in the derivation cohort were compared and shown in Table 2. In the univariate analysis, the “central etiology group” had a higher proportion of males (74.5% vs. 59.3%, p = 0.036), older average age (63.36 vs. 54.85 years, p < 0.01), higher prevalence of hypertension (72.7% vs. 37.0%, p < 0.01), diabetes mellitus (38.2% vs. 12.2%, p < 0.01), and history of stroke (21.8% vs. 3.3%, p < 0.01), as well as higher median ABCD2 scores (4 vs. 3, p < 0.01). The “peripheral etiology group” exhibited higher rates of vertigo symptoms (74.8% vs. 56.4%, p < 0.01), AVS (93.5% vs. 85.5%, p = 0.047), and autonomic symptoms (89.4% vs. 70.9%, p = 0.04). VOG results indicated a significantly higher rate of “peripheral‐type nystagmus patterns” in the peripheral etiology group (98.8% vs. 70.9%, p < 0.01). The peripheral group also showed higher gain asymmetry ratios (0.44 vs. 0.14, p < 0.01) and higher vHIT positivity rates (87.8% vs. 38.2%, p < 0.01).

TABLE 2.

Characteristics of patients in the peripheral and central groups in the derivation cohort.

	Peripheral group (n = 246)	Central group (n = 55)	p
Demographic characteristics
Gender (female)	100 (40.7%)	14 (25.5%)	0.036*
Age	54.85 ± 15.13	63.36 ± 12.39	< 0.001*
Symptomatic features
Vestibular symptom (vertigo)	184 (74.8%)	31 (56.4%)	0.006*
Vestibular vagal symptom (absence)	220 (89.4%)	39 (70.9%)	< 0.001*
≤ 24 h	133 (60.5%)	23 (58.9%)
Mild	18 (13.5%)	6 (26.1%)	0.123
Severe	115 (86.5%)	17 (73.9%)
> 24 h	87 (39.5%)	16 (41.0%)
Mild	17 (19.5%)	3 (18.8%)	0.941
Severe	70 (80.5%)	13 (81.2%)
Cochlear symptom	60 (24.4%)	8 (14.5%)	0.114
Headache symptom	10 (4.1%)	5 (9.1%)	0.121
Duration of last onset	24 (5–72)	24 (3–72)	0.792
≤ 24 h	144 (58.5%)	30 (54.5%)	0.588
> 24 h	102 (41.5%)	25 (45.5%)
Previous attacks	69 (27.6%)	13 (23.6%)	0.545
Vestibular syndrome (acute vestibular syndrome)	230 (93.5%)	47 (85.5%)	0.047*
Medical history
Hypertension	91 (37.0%)	40 (72.7%)	< 0.001*
Diabetes	30 (12.2%)	21 (38.2%)	< 0.001*
Atrial fibrillation	3 (1.2%)	2 (3.6%)	0.205
Stroke	8 (3.3%)	12 (21.8%)	< 0.001*
Physical signs
Beside nystagmus pattern
Null	49 (19.9%)	22 (40.0%)	< 0.001*
Peripheral	189 (76.8%)	22 (40.0%)
Central	8 (3.3%)	11 (20.0%)
Beside HIT
Null	139 (56.5%)	46 (83.6%)	< 0.001*
Negative	13 (5.3%)	4 (7.3%)
Positive	94 (38.2%)	5 (9.1%)
Truncal ataxia
Normal	71 (28.9%)	13 (23.6%)	0.530
Mild	31 (12.6%)	5 (9.1%)
Moderate	45 (18.3%)	9 (16.4%)
Severe	99 (40.2%)	28 (50.9%)
Hearing loss	17 (6.9%)	4 (7.3%)	0.924
ABCD2 score	3 (2–4)	4 (3–5)	< 0.001*
VOG features
VOG nystagmus pattern
Peripheral	243 (98.8%)	39 (70.9%)	< 0.001*
Central	3 (1.2%)	16 (29.1%)
vHIT gain asymmetry rate	0.44 (0.31–0.55)	0.14 (0.08–0.35)	< 0.001*
vHIT a
Negative	30 (12.2%)	34 (61.8%)	< 0.001*
Positive	216 (87.8%)	21 (38.2%)

Abbreviations: HIT, head impulse test; VOG, video oculography.

^a 2 missing.

^* p < 0.05.

3.2. Predictors of Central VS Identified by Multivariable Logistic Regression Analysis

Finally, five independent predictors were included in the multivariable logistic regression model to predict the occurrence of central VS. The following predictors were included (Table 3): vestibular vagal symptoms (OR, 0.233; 95% CI, 0.084–0.648; p = 0.005), previous history of stroke (OR, 15.204; 95% CI, 4.455–51.884; p < 0.001), ABCD₂ score (OR, 1.903; 95% CI, 1.206–3.004; p = 0.006), VOG central nystagmus pattern (OR, 38.377; 95% CI, 8.631–170.644; p < 0.001), and vHIT positive (OR, 0.078; 95% CI, 0.033–0.188; p < 0.001).

TABLE 3.

Multivariable logistic model for isolated vestibular syndrome of central etiology.

Variables	β‐coefficient	SE	Adjusted OR	95% CI	p
Vestibular vagal symptom
Absence	Ref
Presence	−1.457	0.522	0.233	0.084–0.648	0.005
Stroke history
Absence	Ref
Presence	2.722	0.626	15.204	4.455–51.884	< 0.001
ABCD2 score	0.643	0.233	1.903	1.206–3.004	0.006
VOG nystagmus pattern
Peripheral	Ref
Central	3.647	0.761	38.377	8.631–170.644	< 0.001
Video HIT a
Negative	Ref
Positive	−2.547	0.447	0.078	0.033–0.188	< 0.001

Abbreviations: HIT, head impulse test; VOG, video oculography.

^a 2 missing values were entered based on the last available observation.

3.3. Construction and Validation of the SAV₃E Score

Based on the β‐coefficients of the five independent predictors from the multivariable logistic regression model, the corresponding points were assigned to each independent predictor. Therefore, the SAV₃E score was constructed, and its performance was calculated (Table 4 and Tables S1 and S2).

TABLE 4.

SAV₃E score.

Components	SAV3E index
Vestibular vagal symptom
Yes	−1
No	0
Stroke
No	0
Yes	2
ABCD2 score
0–3	0
4–7	2
VOG nystagmus pattern
Peripheral	0
Central	3
Video HIT
Positive	−2
Negative	0
Total index	−3 to 7

Abbreviations: HIT, head impulse test; VOG, video oculography.

The SAV₃E score showed excellent discriminatory performance, with an AUC of 0.910 (95% CI, 0.864–0.956) in the derivation cohort and 0.886 (95% CI, 0.807–0.966) in the validation cohort (Figure 2). Notably, it significantly outperformed both the TriAGe+ model (AUC, 0.706; 95% CI, 0.628–0.785) and STANDING model (AUC, 0.779; 95% CI, 0.864–0.854) in diagnostic accuracy. Calibration analysis revealed good model fit across cohorts, as evidenced by Hosmer‐Lemeshow test results (derivation cohort: p = 0.899, validation cohort: p = 0.789), with calibration curves showing excellent agreement between predicted and observed probabilities (Figure 3). DCA demonstrated good clinical utility of the SAV₃E score across both cohorts (Figure 4). The model exhibited wide effective threshold probability ranges (derivation cohort: 0.01–0.86, validation cohort: 0.01–1.00) with sustained positive net benefits, suggesting excellent clinical applicability for risk stratification.

FIGURE 2.

The receiver operating characteristic curve of the SAV₃E score, along with a comparison to the STANDING and TriAge+ models, is presented to predict central isolated vestibular symptom etiology in both the derivation cohort (A) and the validation cohort (B). SAV₃E score, the area under the curve (AUC): 0.910 (95% CI, 0.864–0.956) in the derivation cohort and 0.886 (95% CI, 0.807–0.966). STANDING score, AUC: 0.779 (95% CI, 0.704–0.854). TriAGe+ score, AUC: 0.706 (95% CI, 0.628–0.785).

FIGURE 3.

Calibration plot of the SAV₃E score to predict the probability of central isolated vestibular symptom etiology in the derivation cohort (A) and the validation cohort (B). The “Ideal” represents the ideal reference line; the “Logistic calibration” represents the model calibration line.

FIGURE 4.

Decision curve analysis of the SAV₃E score to predict the probability of central isolated vestibular symptom etiology in the derivation cohort (A) and the validation cohort (B). The clinical usefulness is the net benefit (y‐axis) of using the score to risk stratify patients relative to two extreme strategies of treating all the patients (the gray line) and treating none of the patients (the black line) across a range of prespecified threshold probabilities (x‐axis). The decision curve demonstrates a positive net benefit of both cohort in nearly all threshold probabilities to predict central isolated vertigo and dizziness etiology.

4. Discussion

In this multicenter retrospective cohort study of 506 patients with isolated VS, we developed and validated the SAV₃E score, a novel clinical prediction model for distinguishing central from peripheral VS etiologies. Our analysis systematically integrated demographic, clinical, and VOG parameters and identified five independent predictors of central VS: absence of vestibular‐vagal symptoms (such as nausea and vomiting) (OR = 0.233), history of stroke (OR = 15.204), elevated ABCD₂ scores (≥ 4) (OR = 1.903), central nystagmus pattern on VOG (OR = 38.377), and positive vHIT results (OR = 0.078). Building on these predictors, the SAV₃E score demonstrated exceptional discriminative performance, achieving AUCs of 0.910 (95% CI, 0.864–0.956) in the derivation cohort and 0.886 (95% CI, 0.807–0.966) in the validation cohort, significantly outperforming established models such as TriAGe+ (AUC = 0.706) and STANDING (AUC = 0.779).

Traditional dizziness etiology prediction models can be classified into two categories: symptom‐based scoring systems and sign‐oriented models [6, 20, 28]. The symptom‐based scoring systems include the TriAGe+ score [28], the DEFENSIVE posterior circulation stroke risk scale [29], and the Sudbury dizziness risk score [30]. The TriAGe+ score integrates demographic characteristics, symptomatology, focal neurological localization signs (e.g., limb weakness and speech impairment), and stroke risk scores, with focal neurological signs carrying the highest weight [28]. The DEFENSIVE scale [29] and Sudbury score [30] also incorporate neurological symptoms and focal signs as core predictive indicators. However, all three tools lack assessment of eye movement signs, which have localizing diagnostic value. This lack of eye movement sign assessment reduces the efficacy in differentiating the etiology of isolated VS.

Compared with symptom‐based scoring systems, the innovations of the SAV₃E model are reflected in the following three aspects. Firstly, the refinement of symptom indicators. Using “vestibulo‐vagal symptoms (containing nausea, vomiting, and diaphoresis)” replaces traditional VS classifications; this replacement avoids variability in symptom interpretation and unreliability in etiological prediction associated with VS types [31]. Secondly, the integration of risk stratification. Combining the ABCD₂ score (sensitivity 61.1%, specificity 62.3% for posterior circulation stroke) with specific predictive indicators improves diagnostic efficacy. When the ABCD₂ score is ≥ 4 points combined with eye movement signs, diagnostic performance significantly improves, reaching a sensitivity of 96.8% and specificity of 98.5% [32]. Moreover, integrating the ABCD₂ score with stroke history or other indicators also elevates posterior circulation stroke prediction specificity to 87.0% [33]. Thirdly, the superiority of ocular motor assessment. VOG parameters account for 50% of the predictive power in the SAV₃E model. The multimodal prediction model demonstrates significantly better discriminative power for isolated VS (AUC = 0.91, 95% CI, 0.864–0.956) compared to the TriAGe+ score (AUC = 0.706, 95% CI, 0.628–0.785).

Sign‐oriented prediction tools, exemplified by HINTS [6] and STANDING [20], share inherent limitations. The HINTS examination exhibited selection bias (with 75.25% stroke prevalence) and diagnostic suspicion bias (51% of cases concurrently presented ocular motor abnormalities and focal neurological signs), both of which compromised its real‐world generalizability in emergency settings [6, 8, 30]. Although STANDING is applicable for all acute dizziness patients and incorporates five sign‐based metrics (e.g., SN, GEN, and HIT), its diagnostic performance remains operator‐dependent [20, 34, 35]. What's more, there were significant missing data for SD sign.

Compared with the sign‐oriented prediction tools, the SAV₃E model overcomes these constraints through three main approaches. First, it replaces subjective SD assessments with quantitative VOG metrics, including nystagmus pattern and vHIT gain, as objective indicators. Second, it transforms bedside nystagmus and HIT evaluations recommended by STANDING into automated VOG and vHIT protocols, thereby resolving the dual challenges of low implementation rates and misjudgment inherent in bedside examinations. Third, it compensates for inadequate risk stratification due to missing SD signs by incorporating complementary clinical predictors.

It is noteworthy that when specific central vestibular structures, such as the medial vestibular nucleus and hypoglossal nucleus, are affected, the resulting eye movement signs may mimic a peripheral nystagmus pattern and positive HIT. Even with the use of VOG combined with vHIT test, clinicians find it difficult to distinguish these “pseudo‐peripheral” eye movement patterns, which are also referred to as “pseudo‐vestibular neuritis/labyrinthitis” [36, 37, 38, 39]. We applied the SAV₃E model on 17 patients with “pseudo‐vestibular neuritis/labyrinthitis” in our cohort. The results showed that this model overcomes the limitations of traditional eye movement techniques by integrating ABCD₂ scores and stroke history, correctly classifying 47.06% of “pseudo‐peripheral” stroke patients (Table S3). This represents a significant improvement over the STANDING interpretation results, which indicated a 0% probability of a central etiology. Among the nine cases of missed diagnosis, seven (77.8%) were due to insufficient total scores caused by the absence of stroke history (with S = 0 indicating no stroke history) combined with an ABCD₂ score of 2. Therefore, this finding indicates the need to adjust the ABCD₂ cutoff value to improve prediction of posterior circulation strokes.

Several limitations warrant consideration. First, the participating centers varied from provincial to county levels, which introduced selection bias. While the efficacy of the model was validated, prospective multicenter studies are still needed to verify its generalizability to external populations, particularly to address confounding factors commonly encountered in emergency departments. Second, due to the inherent limitations of the retrospective cohort design in this study, patients who did not undergo VOG and vHIT were excluded. In addition, patients with indeterminate discharge diagnoses were excluded. This exclusion could lead to an overestimation of the model's predictive performance for vestibular disorders, as such patients may present diagnostic challenges, including symptom variability and diagnostic complexity, that are not captured in the current dataset. Furthermore, there is a lack of dynamic data on important symptoms and signs, such as nystagmus and the vagal symptoms. Third, VOG and vHIT exhibit two main issues related to delayed testing times: (i) the median test delay exceeded 24 h in this cohort, potentially missing the acute‐phase intervention window; (ii) VOG and vHIT were performed significantly later than MRI, which reduces their utility for early risk prediction. Finally, regarding stroke‐induced “pseudo‐vestibular neuritis/labyrinthitis” originating from the vestibular nuclei, nucleus prepositus hypoglossi, or cerebellar flocculus, the model detected only 47.06% of cases despite the integration of clinical parameters. To improve sensitivity, incorporating additional VOG signs (e.g., subjective visual vertical and other relevant metrics) is recommended.

5. Conclusion

This study innovatively developed the multimodal predictive model (SAV₃E) for the etiological diagnosis of isolated VS. By incorporating quantitative VOG parameters, stroke risk stratification parameters, and VS profiles, SAV₃E demonstrated significantly superior discriminatory performance over conventional prediction tools and partially addressed specific diagnostic challenges in VOG related to detecting vestibular neuritis/labyrinthitis mimics.

However, future clinical translation for the SAV₃E model included the following steps. First, developing portable eye‐tracking devices to reduce delays in assessment, combined with AI‐assisted interpretation to minimize operator‐dependent variability. Second, establishing standardized emergency department dizziness pathways to enable electronic SAV₃E parameter collection during triage. Third, incorporating dynamic balance parameters and supplementary VOG metrics (e.g., subjective visual vertical [SVV]) to enhance interpretability. Finally, conducting stepwise validation to evaluate its screening utility.

Author Contributions

Concept and design: Guo Wenting, Shi Tianming, and Guo Shunyuan. Data collection and analysis: Du Liping, Wu Bin, Wang Huiyuan, Yang Yi, Hu Jin, Ye Bijun, Wang Jianer, Liu Shuangsi, Cai Xiaofeng, and Yu Xiang. Drafting of the article: Guo Wenting. Critical revision of the article: Guo Wenting and Shi Tianming. Study supervision. All authors reviewed the manuscript.

Ethics Statement

The protocol of the study was approved by the Ethics Committee of Zhejiang Provincial People's Hospital (No. Zhejiang Provincial People's Hospital Ethics Review 2025 Study No. 037).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Comparison of the predictive performance of different models.

Table S2: SAV₃E score performance measures.

Table S3: Probability of central etiology of patients with “pseudovestibular neuritis/labyrinthitis” in our cohort.

Funding Statement

This work was funded by Zhejiang Medical Science and Technology Project grants 2023KY025 and 2023KY496.

Data Availability Statement

The data supporting the findings of this study are available from the corresponding author on reasonable request.