Sage Journals: Discover world-class research

Abstract

Importance

Clinicians face great challenges in diagnosing dizziness/vertigo disease due to its subjectivity. Currently, there is an absence of machine learning model that could make full use of the information gained from both medical history and physical signs.

Objective

To develop and validate a machine learning model based on medical history and physical signs for dizziness/vertigo disease diagnosis, relieving the burden of diagnosis for clinicians.

Design

A retrospective cohort study.

Setting

Tertiary referral center.

Participants

This study included 1003 patients conformed to the inclusion criteria at the neuro-otologists’ clinics.

Exposures

Thirty-one medical history items, and 9 bedside examination signs recorded by routinely performing a detailed ocular motor examination using video goggles.

Main Outcome Measures

The accuracy, precision, recall, F1 scores, and Matthews’ correlation coefficient of disease diagnosis.

Results

On the collected dataset of 16 categories of dizziness/vertigo diseases, the proposed model achieved an accuracy of 98.11% and an F1 score of 95.43%. The model demonstrated its optimal robustness when tested with datasets containing added noise. Additionally, an analysis of the correlation between medical history and signs was conducted, along with several case studies.

Conclusions

A machine learning-based model was proposed for the diagnosis of dizziness/vertigo, which effectively combined patients’ medical history and signs. In terms of diagnostic accuracy, it outperforms models that rely solely on either medical history or signs for diagnosis.

Relevance

The proposed method can effectively combine the patient’s medical history and physical sign information to make the diagnosis of dizziness/vertigo disease, which has the potential to relieve the burden of diagnosis for clinicians to a certain extent.

Graphical Abstract

Keywords

machine learning diagnosis dizziness/vertigo medical history signs

Key Messages

Previous works remain constrained to isolated analysis of either medical history data or physical signs, lacking integration of both information.

A machine learning-based model is developed to diagnose dizziness/vertigo diseases by analyzing the information from both medical history and physical signs.

In comparison to utilizing either information alone, the combined use of medical history and physical signs demonstrates a substantial improvement in diagnostic performance.

Introduction

According to large population-based studies, the prevalence of dizziness/vertigo among adults varies from 15% to over 20% and rises along with age.¹ Numerous conditions can underlie dizziness/vertigo. A thorough history combined with classic signs make bedside diagnosis possible in most patients with dizziness/vertigo.² Informative history that helps differentiate confusing diagnosis, should include frequency, duration, and triggers of the attacks, previous medical history, and accompanying symptom. Whereas, typical signs may point directly to a specific diagnosis, such as spontaneous, gaze-evoked, head-shaking-evoked, positional-testing evoked nystagmus, and decreased vestibulo-ocular reflex (VOR). Both medical history and body signs are crucial for physicians to make the correct diagnosis, overlooking or over-reliance on either one can easily compromise the accurate assessment.³

Specialists for dizziness/vertigo can reach a preliminary diagnosis of the disease based on detailed medical history and accurate bedside examination results. However, untrained emergency or general practitioners may have a hard time making a diagnosis. In the absence of reliable history or correctly measured physical signs, even experienced doctors may have difficulty making accurate diagnosis. Artificial intelligence algorithms trained on datasets can minimize the effects of missing data in terms of inaccurate medical history or sign judgments; thereby better assisting physicians who are not specialized in dizziness/vertigo with clinical decision-making. When applying machine learning model for dizziness/vertigo diagnosis, most of the methods proposed in the current related literature are limited to either information extracted from the medical history, or data on physical signs obtained from examinations.^4
-6 However, to the best of our knowledge, few works have attempted to make full use of the information gained from both medical history and physical signs. In addition, the current models perform relatively poorly in differentiating the dizziness/vertigo diseases with precision, which rarely includes diagnosis other than the most common 6 to 9 subtypes, such as vestibular migraine (VM), Menière’s disease (MD), benign paroxysmal positional vertigo (BPPV), and vestibular paroxysmia (VP).^7
-9

The aim of this study is to develop and validate a machine learning model based on medical history and physical signs for dizziness/vertigo disease diagnosis, serving as an auxiliary diagnosis tool for clinicians. Specifically, to understand the correlation between medical history and signs, canonical correlation analysis (CCA) was adopted to seek canonical variables, which also establish a shared representation subspace for them at the same time. Inspired by the good performance of ensemble learning, an ensemble prediction model was proposed to collaboratively exploit the information from the original medical history and signs, as well as the potential commonality information between them.

Materials and Methods

Dataset

To develop data-driven machine learning model for dizziness/vertigo disease diagnosis, the dataset used throughout this study was collected from a tertiary referral center. Specifically, in compliance with the medical ethic requirement and privacy regulations, we collected a total of 1003 cases at the neuro-otologists’ clinics. For each case, the medical history and bedside examination signs were recorded, by routinely performing a detailed ocular motor examination using video goggles. In addition, diagnosis of these cases falls into 16 categories, as shown in Table 1, including BPPV, MD, VM, VP, persistent postural perceptual vertigo, hemodynamic orthostatic dizziness/vertigo, severe sudden sensorineural hearing loss with vertigo (excluded from central original). For medical history, 31 typical attributes are extracted to characterize the patients’ baseline health condition. Meanwhile, for physical examination, the nystagmus tests are performed to measure the signs, where 9 associate attributes were extracted. The detailed information about the feature attributes and distribution of demographic characteristics is described in Supplementary Table S1, S2, and Figure S1.

Table 1.

Subtypes of Dizziness/Vertigo Diseases in Dataset.

Index	Disease	No. of cases
1	Acoustic neuroma	8
2	Anterior canal benign paroxysmal positional vertigo	30
3	Autoimmune inner ear disease	24
4	Central vertigo (vascular or nonvascular vertigo)	25
5	Delayed endolymphatic hydrops	40
6	Hemodynamic orthostatic dizziness/vertigo	20
7	Horizontal canal benign paroxysmal positional vertigo	52
8	Hunt syndrome	23
9	Meniere’s disease	198
10	Persistent postural perceptual vertigo	30
11	Posterior canal benign paroxysmal positional vertigo	314
12	Severe sudden sensorineural hearing loss with vertigo	44
13	Vestibular migraine	93
14	Vestibular neuronitis	69
15	Vestibular paroxysmia	17
16	Viral labyrinthitis	16
	Total	1003

Model Overview

As shown in Figure 1a, we develop a machine-learning based model to diagnose dizziness/vertigo diseases by analyzing the information from both medical history and physical signs of patients. Overall, the model mainly consists of 3 parts as follows:

Figure 1.

(a) Overall framework of data-driven prediction and diagnosis of dizziness and vertigo diseases. (b) Schematic representation of feature importance based on CED. CED, case-level explainable paramedical diagnosis model.

Embedding for categorical data is to obtain the numerical type embeddings of the medical history and signs composed of categorical attributions.

CCA attempts to seek canonical variables for correlation measurement between them, which also form a shared representation.

Ensemble prediction is adopted to promote the prediction performance by collaboratively exploiting the information from the original medical history and signs, as well as the potential commonality information between them.

Preliminaries

Let’s first give some notations for use in the following description. Suppose that there are $M$ patients $X = {x_{m}, y_{m}}_{m = 1}^{M}$ composed of $C$ classes in the collected dataset, where $y_{m} \in {1, \dots, C}$ denotes the label of $x_{m}$ . For each patient $m \in {1, \dots, M}$ , he/she is represented as $x_{m} = [x_{m}^{h}, x_{m}^{s}] \in ℝ^{d_{h} + d_{s}}$ where $x_{m}^{h} \in ℝ^{d_{h}}$ and $x_{m}^{s} \in ℝ^{d_{s}}$ are $d_{h}$ -dimensional and $d_{s}$ -dimensional vectors for medical history and sign with categorical attributes, respectively. In addition, for the categorical attribute in medical history, we have $x_{m}^{h} (i) \in A_{i}^{h} = {A_{i, k}^{h}}_{k = 1}^{| A_{i}^{h} |}$ , $i = 1, \dots, d_{h}$ , with $| A_{i}^{h} |$ being the total number of values in attribute $i$ , and the similar definition is for $x_{m}^{s} (j)$ , $j = 1, \dots, d_{s}$ .

CED-Based Embedding for Medical History and Signs

For the given medical history $x^{h} \in ℝ^{d_{h}}$ and signs $x^{s} \in ℝ^{d_{s}}$ , as categorical attributes are used to depict them, it is necessary to obtain their embeddings of the numerical type to facilitate the downstream machine learning models that are limited to dealing with numerical form of data. Just like using support vector machine model for classification, it won’t be workable unless the categorical attributes are encoded into numeric form by one-hot encoding.

In our previous work, we have proposed a decision tree based case-level explainable paramedical diagnosis model (CED).¹⁰ To implement backtracking of the decision path, the bi-side mutual information is exploited in CED to obtain the contribution of each feature to the final decision. By aggregating feature contributions across $T$ trees, we can obtain the feature importance-based embedding of any given input. Taking the medical history $x^{h} \in ℝ^{d_{h}}$ of a patient $x$ as an example, its feature importance based embedding ${\tilde{x}}^{h}$ is given as follows:

{\tilde{x}}^{h} = [\bar{F C} (x^{h} [1]), \bar{F C} (x^{h} [2]) \dots, \bar{F C} (x^{h} [d_{h}])]

(1)

where $\bar{F C} (x^{h} [i]) = \sum_{t = 1}^{T} F C^{(t)} (x^{h} [i])$ with $F C^{(t)} (x^{h} [i])$ denoting the contribution of the $i$ th attribute $x^{h} [i]$ in the $t$ th tree. Similarly, we can obtain the embedding ${\tilde{x}}^{s} \in ℝ^{N_{s}}$ for the patient signs $x^{s}$ by the same way.

In fact, the effects of the embeddings ${\tilde{x}}^{h}$ and ${\tilde{x}}^{s}$ obtained via CED are 2 folds. On the one hand, they convert the categorical attributes of medical history and signs into continuous values, and on the other more important hand, each dimension of them also reflects the importance of corresponding attribute to the final disease diagnosis. In Figure 1b, several cases of the embedding for patient signs consisting of categorical attributes are illustrated.

CCA for Medical History and Signs

As a medical consensus, it is deeply accepted in dizziness/vertigo disease diagnose by clinical doctor that the medical history and signs information from a patient are not independent of each other but take on some kind of correlation to a certain extent; in other word, they generally play a complementary role to each other in making clinical decision. To measure their correlation, the CCA is adopted.¹¹

For multivariate data analysis, CCA aims to seek for multiple pairs of projection vectors that can maximize the correlation between the projections of 2 set of multivariables. Specifically, given the embeddings ${\tilde{X}}^{h} \in ℝ^{M \times d_{h}}$ and ${\tilde{X}}^{s} \in ℝ^{M \times d_{s}}$ of M patients, the canonical correlation between the medical history and signs is given by

Maximize : : T r ({(P_{h})}^{T} \cdot cov ({\tilde{X}}^{h}, {\tilde{X}}^{s}) \cdot P_{s})

\begin{array}{l} S . t . {(P_{h})}^{T} \cdot cov ({\tilde{X}}^{h}, {\tilde{X}}^{h}) \cdot P_{h} \\ = {(P_{s})}^{T} \cdot cov ({\tilde{X}}^{s}, {\tilde{X}}^{s}) \cdot P_{s} = I \end{array}

(2)

where $P_{h} = [p_{h}^{1}, \dots, p_{h}^{d}] \in ℝ^{d_{h} \times d}$ and $P_{s} = [p_{s}^{1}, \dots, p_{s}^{d}] \in ℝ^{d_{s} \times d}$ , $T r (\cdot)$ denotes the trace of a matrix.

Given the canonical correlation subspace supported by $P_{h}$ for ${\tilde{X}}^{h}$ and by $P_{s}$ for ${\tilde{X}}^{s}$ , we use ${\tilde{X}}^{c c a}$ to denote the shared representation between medical history and signs that can be generally formulated as ${\tilde{X}}^{cca} = [{\tilde{X}}^{h} P_{h}, {\tilde{X}}^{s} P_{s}] \in R^{M \times 2 d}$ .

Ensemble Prediction by Voting

As shown in Figure 1a, an ensemble scheme with soft weighting based on gradient boosting decision tree (GBDT) is designed to exploit the information of the original medical history and signs, as well as the potential commonality information between them. Specifically, for a patient $x = [x^{h}, x^{s}]$ , we have

\begin{matrix} E (x^{h}, x^{s}) = w_{h} GBDT (x^{h}) + w_{s} GBDT (x^{s}) \\ + w_{cca} GBDT ({\tilde{x}}^{cca}) \end{matrix}

(3)

where $w_{h}$ , $w_{s}$ , and $w_{cca}$ represent the weights of different classifiers, respectively, which are measured by the accuracy of the respective classifiers on the validation set. $GBDT (x^{h})$ , $GBDT (x^{s})$ , and $GBDT (x^{c c a}) \in ℝ^{C}$ denote the $C$ -dimensional prediction outputs based on the medical history $x^{h}$ , signs $x^{s}$ , and their commonality information, respectively. Hence, the final prediction $\hat{y} \in {1, \dots, C}$ on $x$ can be given as the corresponding index of $E (x^{h}, x^{s})$ obtaining the maximum.

Results

Experiment Setting and Evaluation Metrics

To validate the effectiveness of our proposed model for dizziness/vertigo disease diagnosis, the collected dataset as described in dataset section is divided randomly into training and testing sets in a ratio of 8:2. Meanwhile, the following metrics including accuracy, precision, recall, F1, and Matthews correlation coefficient are adopted for performance evaluation.

Main Results for Dizziness/Vertigo Disease Diagnosis

To evaluate the performances of machine learning models for dizziness/vertigo disease, some classical methods are used as baselines, including logistic regression, multilayer perceptron, categorical boosting, and GBDT. In addition, to make the performance comparison more fair, the random division of the dataset as mentioned above is carried out 5 times, and the average results are reported.

In Table 2, we use “mh,” “ps,” and “mh + ps” to indicate that the medical history, physical signs, and their concatenation, respectively, are used for disease diagnosis. As we can see, the medical history information is more useful for disease diagnosis than physical signs. When taking GBDT as classifier, 15.62% improvement from medical history over physical signs in accuracy is obtained. Meanwhile, in comparison to utilizing either information alone, the combined use of medical history and physical signs demonstrates a substantial improvement in performance across all baseline models. It has shown that our method has achieved the best result in Precision of 96.25%, closely rivaling the best baseline GBDT in other performance metrics, showing the proposed method can effectively integrate medical history and physical signs for diagnosis of dizziness and vertigo diseases.

Table 2.

Results for Dizziness/Vertigo Disease Diagnosis by Using Medical History, Physical Signs, and Their Combination, Respectively. Best and Second-Best Results Are Highlighted in Bold and Italics, Respectively.

Method	Used data	Performance metrics (%)
Method	Used data	Accuracy	Precision	Recall	F1 score	MCC
LR	mh	83.98 (±0.73)	76.28 (±2.98)	75.78 (±3.65)	74.70 (±3.49)	81.94 (±0.89)
	ps	65.87 (±1.16)	34.15 (±3.11)	34.98 (±1.52)	32.58 (±1.40)	60.56 (±1.51)
	mh + ps	97.11 (±0.66)	93.36 (±3.07)	91.97 (±2.56)	92.17 (±2.53)	96.64 (±0.77)
MLP	mh	83.98 (±0.91)	78.04 (±2.98)	77.10 (±3.63)	75.87 (±3.60)	81.97 (±1.10)
	ps	65.67 (±1.30)	32.78 (±1.82)	35.58 (±1.66)	32.75 (±1.59)	60.17 (±1.47)
	mh + ps	97.11 (±0.96)	93.91 (±3.59)	92.94 (±3.33)	93.00 (±3.40)	96.64 (±1.13)
CatBoost	mh	84.68 (±1.06)	77.05 (±4.10)	78.16 (±2.83)	76.88 (±3.61)	82.69 (±1.22)
	ps	67.96 (±1.95)	39.67 (±2.70)	40.19 (±1.96)	38.90 (±2.14)	62.45 (±2.40)
	mh + ps	97.91 (±0.73)	94.38 (±3.08)	92.82 (±2.08)	93.18 (±2.46)	97.57 (±0.85)
GBDT	mh	84.08 (±0.70)	77.91 (±1.71)	77.75 (±1.30)	76.91 (±1.52)	81.95 (±0.90)
	ps	68.46 (±2.63)	42.07 (±4.81)	43.09 (±4.25)	41.70 (±4.31)	63.07 (±3.20)
	mh + ps	98.21 (±1.07)	95.63 (±2.08)	96.56 (±2.81)	95.50 (±2.62)	97.92 (±1.25)
Ours	mh + ps	98.11 (±1.01)	96.25 (±2.41)	95.84 (±2.32)	95.43 (±2.62)	97.81 (±1.18)

Abbreviations: CatBoost, categorical boosting; GBDT, gradient boosting decision tree; LR, logistic regression; MCC, Matthews correlation coefficient; mh, medical history; MLP, multilayer perceptron; ps, physical signs.

To demonstrate the model’s capability in distinguishing rare types, Figure 2 shows the overall confusion matrix of prediction results from 5 repetitive experiments. Through careful observation of the matrix, it can be seen that our model achieves good classification results for rare disease types as well, such as hemodynamic orthostatic dizziness/vertigo (Index 6) and Hunt syndrome (Index 8). This indicates that our model is not only applicable to common vertigo/dizziness diseases but also capable of accurately identifying and classifying rarer types.

Figure 2.

The overall confusion matrix of prediction results.

Robustness of Our Model

For a good machine learning model, it should not only show advantages in overall prediction performance but also have good robustness. In the clinical examination of signs, due to its subjective nature, there will inevitably be incorrect descriptions of certain attributes, which can especially happen to some inexperienced clinicians. To simulate the above situation that may occur in clinical examination of patients, some sort of perturbation to the observed data is carried out.

Compared to medical history, it is more likely to appear as some imprecise judgments in the examination of nystagmus signs. For this reason, we only consider adding some noise to the examination of physical signs. Specifically, we implemented random perturbations on all attributes in the examination results of the nystagmus signs. For each physical sign data of all patients in the test dataset, we randomly replaced the values with one of the other values of the same attribute with a probability of P to simulate the effect of data noise.

To observe the robustness of the decision model under different intensities of interference, we set P = .1 and P = .2, respectively. As we can see from Table 3, the performances of the baseline methods have decreased significantly, whereas our model maintains a notable level of stability and achieves the best result. In real-world scenarios, the inherent limitations of medical conditions and facilities make it inevitable to encounter misjudgments in interpreting patient signs. Our method can effectively reduce the impact of such misjudgment on diagnosis and improve medical convenience and accuracy.

Table 3.

Results for Dizziness/Vertigo Disease Diagnosis After Adding Noise to the Examination of Physical Signs. Best and Second-Best Results Are Highlighted in Bold and Italics, Respectively.

Noise ratio	Method	Performance metrics (%)
Noise ratio	Method	Accuracy	Precision	Recall	F1 score	MCC
P = .1	LR	88.46 (±1.85)	85.68 (±3.01)	82.87 (±2.71)	83.02 (±3.25)	86.71 (±2.06)
	MLP	89.55 (±1.57)	85.19 (±3.31)	84.67 (±4.49)	83.52 (±3.63)	88.03 (±1.78)
	CatBoost	94.13 (±2.19)	90.79 (±2.67)	88.41 (±0.86)	88.80 (±1.37)	93.18 (±2.53)
	GBDT	86.17 (±3.52)	85.04 (±5.81)	84.78 (±6.24)	83.13 (±6.11)	84.34 (±3.83)
	Ours	95.12 (±1.06)	93.15 (±2.05)	91.92 (±2.91)	92.05 (±2.49)	94.33 (±1.23)
P = .2	LR	83.18 (±3.32)	78.78 (±3.94)	75.83 (±3.06)	76.03 (±3.67)	80.80 (±3.65)
	MLP	85.17 (±2.47)	85.06 (±3.16)	83.49 (±2.84)	82.55 (±3.27)	83.21 (±2.72)
	CatBoost	90.25 (±2.64)	85.72 (±3.90)	83.47 (±3.93)	83.56 (±4.12)	88.70 (±3.05)
	GBDT	77.91 (±6.11)	78.84 (±4.56)	71.89 (±7.87)	72.45 (±6.79)	75.12 (±6.73)
	Ours	90.25 (±2.86)	86.30 (±3.76)	85.19 (±1.78)	84.58 (±2.48)	88.71 (±3.35)

Abbreviations: CatBoost, categorical boosting; GBDT, gradient boosting decision tree; LR, logistic regression; MCC, Matthews correlation coefficient; MLP, multi-layer perceptron.

Correlation Analysis Between Medical History and Signs

In order to provide doctors with more information worthy of reference, it is necessary to quantitatively evaluate the correlation between them. As mentioned above, CCA can not only be used to reduce the dimensions of medical history and signs to obtain the shared canonical related variables but also be used to calculate and quantify the correlation of medical history and signs. Concretely, we compute the cosine similarity between the embeddings of medical history and physical signs for each sample and obtain the correlation between the medical history and signs by averaging samples within each disease.

As shown in Figure 3, the correlation between medical history and physical signs for various diseases are consistently positive and exhibit an overall high level. One-sample t-tests (null hypothesis: the correlation in the population is 0) are conducted to evaluate the similarity between the embeddings of medical history and physical signs for each disease. Across all disease cohorts, it can be observed that P < .001. This suggests a strong association between medical history and physical signs within the same disease, facilitating the integration of both feature types for accurate patient diagnosis.

Figure 3.

Correlation analysis between medical history and signs of different subtypes.

Discussion

More than 30 years ago, researchers started to launch studies on the application of artificial intelligence in the field of dizziness/vertigo diagnosis. Since 2019, owing to the pandemic of the COVID-19, artificial intelligence techniques have been increasingly applied to establish diagnosis of dizziness/vertigo in both the emergency room and the remote setting. This change effectively reduced the offline visits, and therefore, significantly lowered the risk of exposure and infection.^12
-16 In 2019, Choi predicted that remote consultation based on artificial intelligence might be the future practice mode that enabled real-time diagnosis of emergency dizziness and dizziness.¹⁷

Machine learning technology has been popularly used in medical record analysis^18
-20 and disease diagnosis^21
-24 throughout the past years. By discovering the underlying features and patterns, machine learning models have shown a great potential for medical-aided diagnosis. As far as dizziness/vertigo disease is concerned, machine learning can be employed for automated detection of nystagmus patterns.^25
-27 Furthermore, both the etiology and treatment options can be determined based on historical data and patients’ clinical manifestations.^28,29 Physicians can, in turn, refer to these prediction results to enhance the accuracy of diagnosis and to optimize the treatment options.

In clinical practice, physicians should integrate information obtained from the medical history with physical signs to establish the diagnosis of dizziness/vertigo. Knowledge of the frequency and duration of dizziness attacks, accompanying symptoms during attacks, patient’s age, precedent attacks of dizziness, past medical history of blood pressure, blood glucose, and lipid-related diseases, past medical history of headache, and hearing loss are all crucial to the diagnosis. Therefore, we constructed a dataset to include the 31 selected attributes concerning the medical history.

All patients underwent a battery of bedside examinations, among which 9 signs were selected as attributes for our model, based on high diagnostic sensitivity and specificity for different vestibular lesions. Specifically, spontaneous nystagmus, gaze-evoked nystagmus, head-shaking nystagmus (HSN), and head-shaking tilt suppression test distinguish central lesions from the peripherals³⁰; head-impulse test; and dynamic visual acuity are typical signs of VOR abnormalities. Furthermore, nystagmus characterizes certain position-related disorders (such as benign paroxysmal positional vertigo (BPPV) and central paroxysmal positional vertigo (CPPV)), including the direction, latency, duration, and attenuation mode of nystagmus. Abnormal HSN results suggested potential vestibular dysfunction, which has the most diagnostic value in distinguishing 2 paroxysmal vestibular lesions with similar clinical symptoms, for example, MD and VM.³¹ Finally, a total of 1003 patients at the neuro-otologists’ clinics was collected.

In this study, we developed a machine learning model based on medical history and physical signs for dizziness/vertigo disease diagnosis, serving as an auxiliary diagnosis tool for clinicians. We conducted experiments using medical history, physical signs, and their combination, respectively. The results showed that diagnosis based on medical history achieved higher accuracy than that based on physical signs alone, while the combination of both achieved superior diagnostic performance compared to using either information source independently. Furthermore, we attempted to add noise to physical signs to simulate the distraction from establishing precise diagnosis in clinical scenario. Despite the background noise, as shown in Table 3, the diagnostic accuracy of our model was as high as 95.12% and 90.25% with random perturbation probability P = .1 and P = .2, respectively. In contrast, baseline methods exhibited marked performance degradation, struggling to adapt to potential physical sign misjudgments in real-world scenarios. It means the proposed model can help those general practitioners or emergency physicians, who are not specialists in dizziness, reach preliminary diagnosis.

As discussed above, diagnostic accuracy ought to be doubted when relying solely on medical history or physical signs in specific clinical scenarios. In contrast, our model is competent to comprehensively integrate both medical history with physical signs, thereby reaching precise diagnosis of the disease. We further demonstrate this through case studies, with specifics detailed in Table 4.

Case 1. A patient diagnosed with central positional vertigo. Due to the metastatic tumor involving his cerebellum, this patient suffered clinical symptoms resembling closely to that of BPPV. It is hard to avoid misdiagnosis if medical history was the only evidence to be considered. Signs including consistent spontaneous downbeat nystagmus, directional changing gaze-evoked nystagmus, and atypical direction of nystagmus evoked by positional tests all pointed to the central origin.

Case 2. The clinical manifestations of different subtypes of BPPV are the same, characterized by spinning sensation provoked by changes in head position. However, for different subtypes of BPPV there are different liberatory maneuvers. In case 2, diagnosis of posterior BPPV can only be confirmed via direction, latency, and duration of the patient’s nystagmus, which allowed for the administration of appropriate treatment.

Case 3. For this case, it indicates the importance of medical history to reaching an accurate diagnosis. MD and VM, the 2 most common diseases of episodic vestibular syndrome, might share similar physical signs, such as horizontal HSN. However, accompanying symptoms might offer insights into differential diagnosis, for example, ear muffling with fluctuating hearing loss suggests MD, while either headache or the complication of photophobia and phonophobia indicates VM.

Although the proposed method has achieved promising results in dizziness/vertigo disease diagnosis, there are several limitations in this study. First, its retrospective, single-center design may inherently restrict the model’s generalizability. In the future, we plan to establish multicenter collaborations to conduct external validation of our model in a larger cohort. Moreover, our current work lacks integration with domain knowledge in medicine. In future work, we intend to construct a specialized medical knowledge graph, leveraging structured medical expertise to enhance the representations of medical history and physical signs. More broadly, medical artificial intelligence faces some ethical challenges. First, regarding privacy and security, model training relies on large-scale datasets that contain sensitive patient information and health records. Unauthorized disclosure or misuse of such data could compromise individual privacy and pose significant risks. Second, regarding the reliability of the results, although the model achieves high accuracy, the decision-making process by which it generates its final outputs remains opaque and unintuitive, which may engender patient concern and distrust. Additionally, biases embedded in data collection or algorithmic design may lead to unfair outcomes for specific populations, potentially leading to erroneous decisions and even undermining social justice and stability.

Table 4.

Case Studies of Employing Medical History and Physical Signs for Predicting Disease Diagnosis.

Case	Diagnosis	Prediction
Case	Diagnosis	GBDT-mh	GBDT-ps	Ours
1	Central vertigo (central positional vertigo)	Posterior canal BPPV ×	Central vertigo (central positional vertigo) ✓	Central vertigo (central positional vertigo) ✓
2	Anterior canal BPPV	Posterior canal BPPV ×	Anterior canal BPPV ✓	Anterior canal BPPV ✓
3	Vestibular migraine	Vestibular migraine ✓	Meniere’s disease ×	Vestibular migraine ✓

Abbreviations: BPPV, benign paroxysmal positional vertigo; GBDT-mh, gradient boosting decision tree solely based on medical history; GBDT-ps, gradient boosting decision tree solely based on physical signs.

Conclusion

In this article, we propose a machine learning model to realize accurate diagnosis of dizziness/vertigo disease, thereby providing auxiliary diagnosis. The proposed model collaboratively exploits the information from both the original medical history and physical signs, as well as the potential commonality information between them. The experimental results indicated that our model achieved excellent performance and also demonstrated certain robustness. It means that the developed model can relieve the burden of diagnosis for clinicians to a certain extent.

Supplemental Material

sj-docx-1-ohn-10.1177_19160216251375034 – Supplemental material for Machine Learning-Based Dizziness/Vertigo Disease Diagnosis by Combining Medical History and Signs

Supplemental material, sj-docx-1-ohn-10.1177_19160216251375034 for Machine Learning-Based Dizziness/Vertigo Disease Diagnosis by Combining Medical History and Signs by Yiwen Zhao, Xumeng Tian, Haiyan Wu, Muhao Xu, Ruizhe Yang, Jinlin Xiao and Zhenfeng Zhu in Journal of Otolaryngology - Head & Neck Surgery

Footnotes

Author Contributions

Y.Z. contributed to software, investigation, methodology, writing—original draft. X.T. contributed to investigation, methodology, visualization, writing—original draft. H.W. contributed to conceptualization, data curation, project administration, writing—review and editing. M.X. contributed to formal analysis, validation. R.Y. contributed to writing—review and editing. J.X. contributed to software, visualization. Z.Z. contributed to supervision, formal analysis, writing—review and editing.

Data Availability Statement

The dataset used during the current study are available from the corresponding author on reasonable request.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was in part by National High Level Hospital Clinical Research Funding (2022-PUMCH-C-041), Beijing Natural Science Foundation (7222313), and Non-profit Central Research Institute Fund of Chinese Academy of Medical Sciences (2023-RW320-03).

Ethical Considerations

This study was exempted from the Institutional Review Board review by the Medical Ethics Committee of Peking Union Medical College Hospital due to the retrospective design of the study (No I-22PJ211).

Consent for Publication

Not applicable.

ORCID iDs

Muhao Xu

Zhenfeng Zhu

Supplemental Material

Additional supporting information is available in the online version of the article.

References

Neuhauser

HK.

The epidemiology of dizziness and vertigo. Handb Clin Neurol. 2016;137:67-82. doi:10.1016/B978-0-444-63437-5.00005-4

Newman-Toker

DE.

Symptoms and signs of neuro-otologic disorders. Continuum. 2012;18(5):1016-1040. doi:10.1212/01.CON.0000421618.33654.8a

Edlow

Gurley

Newman-Toker

DE.

A new diagnostic approach to the adult patient with acute dizziness. J Emerg Med. 2018;54(4):469-483. doi:10.1016/j.jemermed.2017.12.024

Neuhauser

Radtke

von Brevern

Lezius

Feldmann

Lempert

Burden of dizziness and vertigo in the community. Arch Intern Med. 2008;168(19):2118-2124. doi:10.1001/archinte.168.19.2118

Kim

Jang

Kim

, et al Diagnosis of acute central dizziness with simple clinical information using machine learning. Front Neurol. 2021;12:691057. doi:10.3389/fneur.2021.691057

Anh

Takakura

Asai

Ueda

Shojaku

Application of machine learning in the diagnosis of vestibular disease. Sci Rep. 2022;12(1):20805. doi:10.1038/s41598-022-24979-9

Groezinger

Huppert

Strobl

Grill

Development and validation of a classification algorithm to diagnose and differentiate spontaneous episodic vertigo syndromes: results from the DizzyReg patient registry. J Neurol. 2020;267:160-167. doi:10.1007/s00415-020-10061-9

Filippopulos

Strobl

Belanovic

, et al Validation of a comprehensive diagnostic algorithm for patients with acute vertigo and dizziness. Eur J Neurol. 2022;29(10):3092-3101. doi:10.1111/ene.15448

Deng

, et al A Questionnaire-based ensemble learning model to predict the diagnosis of vertigo: model development and validation study. J Med Internet Res. 2022;24(8):e34126. doi:10.2196/34126

10.

Guo

Yang

, et al CED: a case-level explainable paramedical diagnosis via AdaGBDT. Comput Biol Med. 2023;153:106500. doi:10.1016/j.compbiomed.2022.106500

11.

Hotelling

. Relations between two sets of variates. In: Kotz

Johnson

, eds. Breakthroughs in Statistics. vol. 14. Springer; 1992:162-190.

12.

Green

Pogson

Otero-Millan

, et al Opinion and special articles. Remote evaluation of acute vertigo: strategies and technological considerations. Neurology. 2021;96(1):34-38. doi:10.1212/WNL.0000000000010980

13.

Murdin

Saman

Rea

The remote neuro-otology assessment—managing dizziness in the coronavirus disease 2019 era. J Laryngol Otol. 2020;134(12):1120-1122. doi:10.1017/S0022215120002273

14.

Fancello

Bianchini

Iannella

, et al

The evaluation of vertigo in children: which role for parental-camera recording and telemedicine?

Ear Nose Throat J. 2022;103(9):NP578-NP580. doi:10.1177/01455613221074739

15.

Zeleňák

Krajina

Meyer

, et al How to improve the management of acute ischemic stroke by modern technologies, artificial intelligence, and new treatment methods. Life. 2021;11(6):488. doi:10.3390/life11060488

16.

Müller-Barna

Hubert

Bergner

, et al TeleVertigo: diagnosing stroke in acute dizziness: a telemedicine-supported approach. Stroke. 2019;50(11):3293-3298. doi:10.1161/STROKEAHA.119.026505

17.

Choi

Kim

JS.

Vascular vertigo: updates. J Neurol. 2019;266(8):1835-1843. doi:10.1007/s00415-018-9040-3

18.

Karadaghy

Shew

New

Bur

AM.

Development and assessment of a machine learning model to help predict survival among patients with oral squamous cell carcinoma. JAMA Otolaryngol Head Neck Surg. 2019;145(12):1115-1120. doi:10.1001/jamaoto.2019.0981

19.

Chen

You

Yang

, et al Machine learning is an effective method to predict the 90-day prognosis of patients with transient ischemic attack and minor stroke. BMC Med Res Methodol. 2022;22(1):195. doi:10.1186/s12874-022-01672-z

20.

Zhu

, et al Cooperative dual medical ontology representation learning for clinical assisted decision-making. Comput Biol Med. 2023;163:107138. doi:10.1016/j.compbiomed.2023.107138

21.

Ahmadi

Vivar

Frei

, et al Towards computerized diagnosis of neurological stance disorders: data mining and machine learning of posturography and sway. J Neurol. 2019;266:108-117. doi:10.1007/s00415-019-09458-y

22.

Wilson

Ali

Kovatch

Smith

Hoff

PT.

Machine learning diagnosis of peritonsillar abscess. Otolaryngol Head Neck Surg. 2019;161(5):796-799. doi:10.1177/0194599819868178

23.

Zeng

Kang

Chen

, et al A deep learning approach to predict conductive hearing loss in patients with otitis media with effusion using otoscopic images. JAMA Otolaryngol Head Neck Surg. 2022;148(7):612-620. doi:10.1001/jamaoto.2022.0900

24.

Zheng

Zhu

Liu

, et al Multi-modal graph learning for disease prediction. IEEE Trans Med Imaging. 2022;41(9):2207-2216. doi:10.1109/TMI.2022.3159264

25.

Zhang

Liu

, et al Deep learning based torsional nystagmus detection for dizziness and vertigo diagnosis. Biomed Signal Process Control. 2021;68:102616. doi:10.1016/j.bspc.2021.102616

26.

Rastall

Green

Deep learning in acute vertigo diagnosis. J Neurol Sci. 2022;443:120454. doi:10.1016/j.jns.2022.120454

27.

Lee

Han

Seo

Yang

A nystagmus extraction system using artificial intelligence for video-nystagmography. Sci Rep. 2023;13(1):11975. doi:10.1038/s41598-023-39104-7

28.

Ahmadi

Vivar

Navab

, et al Modern machine-learning can support diagnostic differentiation of central and peripheral acute vestibular disorders. J Neurol. 2020;267:143-152. doi:10.1007/s00415-020-09931-z

29.

Formeister

Baum

Sharon

JD.

Supervised machine learning models for classifying common causes of dizziness. Am J Otolaryngol. 2022;43(3):103402. doi:10.1016/j.amjoto.2022.103402

30.

Zuma

Maia

Cal

D’Albora

Carmona

Schubert

MC.

Head-shaking tilt suppression: a clinical test to discern central from peripheral causes of vertigo. J Neurol. 2017;264(6):1264-1270. doi:10.1007/s00415-017-8524-x

31.

Jeong

Kim

Koo

Kim

JS.

Vestibular dysfunction in migraine: effects of associated vertigo and motion sickness. J Neurol. 2010;257(6):905-912. doi:10.1007/s00415-009-5435-5

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

0.05 MB