With a lifetime prevalence between 20 and 30% (1), vertigo and dizziness (VaD) belong to the most common complaints in primary care and emergency departments (2, 3). VaD have an annual prevalence of 9% in medical claim databases (4) and a high impact on daily life (5, 6).

In most cases, the diagnosis and treatment of VaD is straightforward (1, 7, 8). Inappropriate or delayed management of VaD, however, may contribute to chronic symptoms, increase disability, and cause secondary psychosomatic disorders (9–11).

While patients with VaD are mostly processed in primary care, a recent systematic review found considerable inconsistencies in the management of dizzy patients in this setting (12). Experts claim that taking a structured medical history is the key to make basic triage, namely, to decide if the patient can be managed in primary care, needs referral to a specialist, or, in the rare case of a life-threatening etiology of VaD (e.g., stroke), even needs emergency care. Concepts for a structured history taking such as the “Five Keys” (13) proposed that a limited number of symptom characteristics allows to successfully differentiate a majority of all VaD cases. Commonly recommended questions for rational history taking are whether VaD the complaints are paroxysmal or permanent, short or long lasting, spontaneous or triggered, of stereotypical presentation, direction-specific, or accompanied by neurological or otological symptoms (13, 14).

However, it has been argued that successful medical reasoning is often based on implicit experience and “gut feeling” that supplements explicit structured knowledge (15). Intuitive elements in the diagnostic process seem to play an important role in diagnostic reasoning (16). For example, the physician's feeling that “there was something wrong” turned out to predict serious infections in children (17). Personal preferences play an important role not only in taking medical history but also for assessing diagnostic thresholds and how to interpret them (18), probably as a function of experience and specialization. In this case, history taking in VaD may be less evidence-based than proposed.

In the current study, we validated a set of questions proposed by neurootological experts by a data-driven approach and hypothesized that a limited number of key characteristics will facilitate the differentiation of the most prevalent vestibular disorders with sufficient validity. To identify these key characteristics, several different statistical approaches are possible. In a previous study, we could show that machine learning methods may display good prediction accuracy but do not yield information about the causal pathways leading to good prediction (19). However, there are methods such as classification and regression trees (20) and their more recently developed methodological refinements (21) to rank characteristics according to their importance.

The aim of this study is to identify a set of key variables that are based on both expert experience and empirical knowledge, can be easily collected in clinical practice, and can be used as indicators for correct diagnosis of the most common vestibular disorders.

To assemble an a priori collection of variables that are meaningful according to expert opinion, we first conducted a worldwide online survey among experts in the field about relevant themes and aspects of history taking in vestibular disease. We then used qualitative content analysis to clarify and structure these aspects. Thirdly, resulting contents were linked to variables from a specialized clinical patient registry that contained verified diagnoses according to current diagnostic guidelines as a gold standard. Fourthly, we used classification trees to analyze if these variables with apparent face validity also had statistical predictive validity.

In total, 21 experts from 16 different countries took part in the online survey. The experts worked in a total of 19 centers treating an average of 1,000 patients per year (median: 700, range: 40–4,000). The participants reported an average of 23 years of clinical experience (median: 25, range: 4–42). A total of 152 different statements were reported.

Content analysis yielded nine main categories and 39 subcategories, which are shown in Table 1.

A total of 98 variables contained in the DizzyReg could be linked to one of the categories. Ten categories were not represented in the registry, e.g., alcohol or pressure changes as a trigger. A complete description of the linking results can be found in the electronic appendix (Supplementary Table 1).

We included a total of 1,066 patients with a mean age of 51.1 years (standard deviation, SD = 15.3), 56% of whom were female. Functional dizziness was the most frequent diagnosis (32.5%), followed by vestibular migraine (20.2%) and Menière's disease (13.3%). A total of 47% of patients had vertigo or dizziness for <2 years (for more details, see Table 2).

Eight variables were found to be indicative for vertigo and dizziness diagnoses: attack duration, rotational vertigo, hearing problems, turning in bed as a trigger, doing sport or heavy household chores as a trigger, age, having problems with walking in the dark, and vomiting. The resulting tree is shown in Figure 1.

To give an example for interpretation: two paths in the tree could identify 56.7% of patients with BPPV correctly. In the first path, a short attack duration (<2 min) and turning in bed as a trigger lead to a correct classification of 55 patients with BPPV. In the second path, a longer attack duration, no hearing problems, age >60, and turning in bed as a trigger lead to 21 patients being correctly classified. In summary, of the 134 patients with BPPV, 76 (56.7%) were assigned to the correct classification of BPPV. The overall accuracy of the algorithm for all diagnoses as a multi-class problem was 42.2%. Further details of the diagnostic properties can be found in Table 3.

The variable importance of the 20 variables that contributed most to the classification was determined by random forest analysis and is shown in Table 4. The variable with the most influence on the performance of the algorithm was vomiting, followed by age and hearing problems.

This study was able to identify a set of eight key questions that can help to differentiate seven common vestibular diagnoses. The key questions comprised attack duration, rotational vertigo, hearing problems, turning in bed as a trigger, doing sport or heavy household chores as a trigger, age, having problems with walking in the dark, and vomiting.

The negative predictive values were higher than the positive predictive values, indicating that it was mostly easier to exclude a diagnosis than to confirm it.

Using expert opinion and a statistical classification approach, we arrived at combinations of symptoms with high face validity. Positive predictive value in our study was highest for functional dizziness. This is not surprising because FD is characterized by the combination of typical symptoms and the absence of others (38). Thus, the sequence of longer attack duration, no hearing problems, younger age, no vegetative symptoms, and a presentation as dizziness rather than rotational vertigo indicated FD, which is in line with the approach presented by Dieterich et al. (54, 55). On the other hand, FD was also frequently present in other nodes (between 13 and 29%). This finding may be explained by the relatively high prevalence of FD in this sample. Furthermore, patients with FD report a multitude of uncharacteristic symptoms fluctuating in time and intensity, triggered by various situations (55). Furthermore, FD often manifests as a comorbidity to or consequence of different organic vestibular disorders, most commonly VM, BPPV, and MD (56).

BPPV was likewise characterized in our study by a short duration of attacks and the movement of body and head in the horizontal plane, which is in line with the typical clinical presentation (43, 57). These two questions identified more than half of BPPV patients. Head movement while turning in bed was also indicative for an older group of patients with BPPV who had longer attacks. This finding also aligns with literature (58).

In addition, the characterization of MD, BVP, and UVP in our study is in accordance with the clinical key features described previously (39, 59, 60). In our study, MD patients had a longer attack duration and hearing loss. BVP was characterized by symptoms aggravating in darkness, permanent dizziness, and higher age. Increase of symptoms in darkness and older age are highly characteristic for BVP (61). UVP characteristics were a longer duration of symptoms, lack of hearing loss, younger age (<60 years), vomiting, and aggravation by sports or heavy household chores. Although UVP can occur in all age groups and is thus not considered to be typical for younger adults, a peak between the ages of 30 and 50 was suggested (62).

Vestibular migraine was difficult to classify in our study. The clearest differentiation was age, which is confirmed by the finding that VM typically manifests in younger adults (63) without typical triggers (34). Interestingly, we were not able to confirm headache as a typical feature of VM. It has been shown that about 30% of patients with VM do not report headache associated to vertigo attacks (64), and <50% of patients report the simultaneous presence of headache and vertigo during attacks (65).

A specific objective of the present study was to train a machine learning model for classification that is transparent, easy to use in daily clinical practice, and easy to understand. In the past, several different approaches have been used to help classify the underlying pathologies of vertigo and dizziness. A complex method like deep neural networks (DNN), which was applied to vestibular disorders with promising results (19, 66), is difficult to directly transfer to a real-world clinical setting since DNN does not provide transparency on how the classification decision was made. Similarly, applying support vector machine (SVM) is accompanied by a lack of procedural transparency but comes with a high power regarding accuracy. A study on classifying unilateral vestibulopathy using SVM yielded an accuracy of 76% (67). Another study was able to differentiate the peripheral and central causes of acute vestibular disorders with a high accuracy using modern machine learning methods (68). Recent research aimed to alleviate the drawback of procedural opacity and develop explainable artificial intelligence (69), but their work was based on pattern recognition and has not been applied to data sets with numeric, ordinal, or categorical data yet. Thus, further research is required to identify if such an approach might be appropriate for a clinical decision-making system in specialized areas like vestibular disorders.

Studies using transparent learning methods, like CART, are scarce. A study from 2000 also applied classification trees for differentiation of vestibular disorders (70). They identified hearing loss, duration of the disease, frequency of attacks, severity of rotational vertigo, onset and type of hearing loss, and occurrence of head injury at onset of vertigo as important variables for diagnostic classification. These variables are similar to the ones presented here, although some variables were not surveyed in the present study (e.g., association to trauma). Another study used boosted decision trees to identify two different feature sets, one for general practitioners and one for experts (71). All these studies reported a higher accuracy than our study, as they used a one-vs.-all classification approach, which results in better accuracy but is less precise than our approach. To put this into context, a one-vs.-all approach has an implicit minimal accuracy of 50% as the classification problem is reduced to a dichotomous choice. In contrast to this, our algorithm aims at distinguishing between seven different vertigo syndromes simultaneously, yielding a minimal accuracy of 1/7 = 14%. Thus, the overall accuracy of 42.2% is a notable improvement.

The data analyzed in this study is subject to the following licenses/restrictions: an application form has to be sent to the Scientific Committee of the German Center for Vertigo and Balance Disorders. Requests to access these datasets should be directed to ralf.strobl@med.uni-muenchen.de.

The studies involving human participants were reviewed and approved by Ethic committee of the medical faculty, LMU Munich, Munich, Germany. The patients/participants provided their written informed consent to participate in this study.

RS contributed to the drafting/revising of the manuscript for content, including writing, study concept/design, interpretation of data, and acquisition of data. MG contributed to statistical analysis and revision of the manuscript for content, including writing, and study concept. AZ contributed to the interpretation of data and revision of the manuscript for content. DH contributed to the development of the concept and revision of the manuscript for content. FF contributed to the interpretation of data, revision of the manuscript for content, and medical writing. EG contributed to the drafting/revision the manuscript for content, including medical writing and study concept/design. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.