Machine Learning‐Based Mobile Application for Predicting Posterior Canal Benign Paroxysmal Positional Vertigo

ABSTRACT

Objective

This study investigated the predictability of the Posterior Canal Being Paroxysmal Positional Vertigo (PC‐BPPV) using vertigo/dizziness features and medical history in machine learning. Secondly, this study aimed to develop a mobile application using the model that predicts PC‐BPPV with the highest accuracy rate.

Methods

This study retrospectively analyzed the medical records of patients who presented to the Audiology and Balance Clinic with complaints of dizziness or vertigo between 04/01/2021 and 09/16/2023. Patients' diagnoses, demographic information, medical history, and dizziness/vertigo characteristics were used in 8 different machine learning models. A mobile application was developed with the model with the highest accuracy.

Results

The study included data from 280 patients. Age, symptom onset time, duration of symptoms, dizziness type, triggering factors, and auditory symptom status were the distinguishing factors for PC‐BPPV. Using these features, the Random Forest algorithm predicted PC‐BPPV with 96.43% accuracy. The accuracy rates of other algorithms were between 89.28% and 94.64%.

Conclusion

Dizziness/vertigo characteristics and medical history can be effectively utilized in machine learning to predict BPPV with high accuracy. The mobile application developed using this algorithm underscores the potential of artificial intelligence platforms to contribute to vestibular science in the telemedicine field.

Level of Evidence

Level 2.

1. Introduction

Benign Paroxysmal Positional Vertigo (BPPV) is the most common type of vertigo, caused by the displacement of otoconia from the utricle into the semicircular canals [1]. This displacement can occur secondary to conditions such as Meniere's disease or head trauma, or it may be idiopathic. Dislodged otoconia either move freely within the canals (canalolithiasis) or adhere to the cupula (cupulolithiasis) leading to abnormal stimulation of the vestibulo‐ocular reflex during head movements [2]. These abnormal signals cause brief but severe vertigo attacks triggered specifically by head motion, with few or no symptoms at rest.

BPPV predominantly affects older women, peaking in the 60s, with a female‐to‐male ratio of 2.4:1 [3]. Posterior canal (PC) BPPV is the most frequent subtype, accounting for approximately 90% of cases [4]. Symptoms often occur when lying down, turning in bed, or rising, owing to the orientation of the posterior semicircular canal. Diagnosis typically relies on observing characteristic nystagmus during the Dix–Hallpike maneuver [5]. Effective canalith repositioning maneuvers lead to recovery in about 80% of cases after a single attempt and in 92% with repeated attempts [6, 7]. Even without intervention, spontaneous remission can occur within approximately 17 days in one‐fifth of patients [8, 9].

Machine learning has increasingly been applied across various medical fields to enhance diagnostic accuracy by extracting meaningful patterns from complex clinical data [10]. These models are typically trained on large data sets associated with diseases or prognostic outcomes, enabling systems to learn the underlying relationships between clinical variables and specific diagnostic or prognostic endpoints [10]. Subsequently, the weighted models are expected to predict outcomes based on new input variables. Similarly, clinical features that distinguish PC‐BPPV from other causes of dizziness can be integrated into machine learning models to optimize diagnostic performance. However, their application in the diagnosis of PC‐BPPV remains limited [11, 12]. Although PC‐BPPV is typically diagnosed through clinical history and positional tests, diagnostic challenges persist, particularly among non‐specialist clinicians in primary care or emergency settings [13]. Delayed diagnosis can increase the risk of falls, contribute to fear of falling, and ultimately diminish quality of life. Developing machine learning‐based models capable of predicting PC‐BPPV could facilitate the early recognition and management of a broader patient population, particularly through telemedicine applications.

The aim of this study is to investigate the predictability of PC‐BPPV by utilizing vertigo/dizziness characteristics and medical history in machine learning. Additionally, our study aims to develop a mobile application using the model that predicts PC‐BPPV with the highest accuracy rate. This mobile application is intended to provide users with preliminary information about the likelihood of having BPPV before consulting a healthcare facility.

2. Methods

Ethical approval for this retrospective study was obtained from the Karabük University Non‐Interventional Ethics Committee (Decision No: 2024/1650). The requirement for informed consent was waived due to the use of anonymized data. All procedures were conducted in accordance with the principles of the Declaration of Helsinki and the guidelines for Good Clinical Practice. For this study, data from patients with dizziness or vertigo who visited the Audiology and Balance Clinic between 01.04.2021 and 16.09.2023 were examined. Patients' medical history, videonystagmography (gaze, oculomotor tests, and spontaneous nystagmus test) positional tests (Dix Hallpike, supine roll and deep head hanging), caloric test and video head impulse test (vHIT) results were evaluated. The patients' ages, genders, diseases, and dizziness characteristics (symptom duration, attack duration, dizziness type, auditory symptoms (tinnitus, hearing loss and fullness) and triggering factors) were noted for use in machine learning. The following criteria were applied in determining the final form of patient data to be used in the study:

3. Inclusion Criteria

1. Patients aged 10 years and older who reported complaints of vertigo or dizziness

2. Patients with complete records regarding age, gender, comorbidities, medical history, vestibular test findings, and vertigo/dizziness characteristics.

4. Exclusion Criteria

1. Subtypes of BPPV for which there is an insufficient number of cases available for machine learning analysis.

2. Patients with unclear diagnostic categorization despite the presence of positional nystagmus, or those with vestibular disorders overlapping with BPPV (e.g., Meniere's disease and BPPV)

3. Patients with acute central pathologies such as cerebellar infarction, brainstem stroke, and transient ischemic attack.

Statistical analysis was performed between the groups to determine the features (input) used in machine learning algorithms. Data from patients meeting the inclusion criteria were used in 8 different machine learning algorithms. The heat map was used to determine the relationship between the determined features and the results.

4.1. Statistical Analysis and Power Analysis

To determine the minimum sample size required for reliable validation of the binary classification model, we applied the Sample Size Analysis for Machine‐Learning (SSAML) framework as proposed by Goldenholz et al. [14] Bootstrapped estimates were used to compute the relative width of the confidence interval (RWD), estimation bias (BIAS), and coverage probability (COVP) for 95% confidence intervals. A sample size was considered adequate if it met the criteria of RWD < 0.5, |BIAS| < 0.05, and COVP > 0.95 for all three metrics. Based on this analysis, the minimum required sample size for our dataset was determined to be 120 subjects.

All statistical analyses were performed using IBM SPSS 21 software (SPSS, Chicago, IL, USA). The normality of the distribution was assessed using the Shapiro–Wilk test. Normally distributed data were presented as mean ± standard deviation (ss), and non‐normally distributed data as median (min‐max). For comparisons between two groups, the Mann–Whitney U test was used for non‐normally distributed continuous variables, while categorical variables were analyzed using the Chi‐square test or Fisher's exact test where appropriate. Given the multiple comparisons made, the Benjamini–Hochberg procedure was applied to adjust p‐values and control the false discovery rate, with a significance threshold set at 0.05.

4.2. Machine Learning Algorithms

Python version 3.7 was used for machine learning algorithms. Eight different machine learning algorithms were used in our study. These algorithms were K‐Nearest Neighbor, Decision Trees, Random Forest, Support Vector Machine, Logistic Regression, Naive Bayes, XGBoost, and artificial neural networks. Models were trained with 80% of the data and tested with 20%. The hyperparameters were determined using the Grid Search method. The hyperparameters used while training the models are presented in Table 1. A confusion matrix for the testing phase was created. According to the confusion matrix, the models' precision, F1‐score, accuracy, recall, and area under the Receiver Operating Characteristic Curve (AUC‐ROC) were calculated. For model success, accuracy and AUC‐ROC values were considered, respectively. The formulas for the performance metrics and the machine‐learning algorithms used are presented below.

$$\mathit{\text{Accuracy}}=\frac{\mathit{\text{True Positive}}+\mathit{\text{True Negative}}}{\mathit{\text{True Positive}}+\mathit{\text{False Positive}}+\mathit{\text{True Negative}}+\mathit{\text{False Negative}}}$$

$$\mathit{\text{Precision}}=\frac{\mathit{\text{True Positive}}}{\mathit{\text{True Positive}}+\mathit{\text{False Positive}}}$$

$$\mathit{\text{Recall}}=\frac{\mathit{\text{True Positive}}}{\mathit{\text{True Positive}}+\mathit{\text{False Negative}}}$$

$$F1-\mathit{\text{score}}=2\times \frac{\mathit{\text{Precision}}\times \mathit{\text{Recall}}}{\mathit{\text{Precision}}+\mathit{\text{Recall}}}$$

TABLE 1.

Hyperparameters used in the algorithms.

Algorithms	Hyperparameters
K‐ Nearest Neighbor	n_neighbors = 27, metric = ‘minkowski’, p = 1
Naive Bayes	Default
Decision Tree	max_depth = 5 criterion = ‘entropy’
Random Forest	n_estimators = 10, max_depth = 4 criterion = ‘entropy’
Support Vector Machines	kernel = ‘linear’, C = 0.1
Logistic Regression	Default
XGBoost	max_depth = 10 eval_metric = ‘mlogloss’ learning_rate = 0.01
Artificial Neural Networks	batch_size = 2
	Total 4 layers, Input layer neuron count: 8 1. Hidden layer neuron count: 6 2. Hidden layer neuron count: 4 Output layer neuron count:1, epoch: 1000, loss = binary_crossentropy, optimizer = ‘adam’, learning_rate = 0.001

4.3. K‐Nearest Neighbor

K‐Nearest Neighbor is a non‐parametric algorithm used for classifying data points based on the majority class of their k nearest neighbors. When a new data point is introduced, the algorithm identifies the k nearest neighbors from the training set and uses their class labels to predict the class of the unknown data point. The Euclidean distance metric is commonly used for this purpose [15].

4.4. Decision Trees

Decision Trees are an algorithm that builds a tree‐like model of decisions, where each node represents a feature‐based decision rule, and branches represent the outcomes of those decisions. The tree starts from the root node, splits the data at each internal node based on a test condition, and leads to leaf nodes where predicted values or classes are assigned. While the algorithm is interpretable and easy to visualize, it tends to overfit, especially with deep trees [15].

4.5. Random Forest

This algorithm is an ensemble learning model that creates multiple decision trees by randomly selecting subsets of the dataset and features, and then combines these trees into a forest structure. This method generates multiple diverse trees, each making different decisions, and does not perform pruning on the trees, which helps reduce the problem of overfitting [15].

4.6. Support Vector Machine

This algorithm aims to find a separating hyperplane that best divides the inputs to achieve the desired result. The main goal of the algorithm is to ensure correct separation and maximize the distance between the hyperplane and the inputs that have opposing effects on the outcome. It is effective in high‐dimensional spaces and is robust to outliers [15].

4.7. Logistic Regression

Logistic Regression is a linear model primarily used for binary classification problems. This model uses the logistic (sigmoid) function to predict the probability of each class. Logistic Regression is widely preferred due to its simplicity and interpretability, making it easy to apply and understand. However, since this model is based solely on linear relationships, it may struggle to handle nonlinear relationships, which can limit the model's performance [15].

4.8. XGBoost

The XGBoost algorithm is a tree‐based model that uses the gradient boosting approach. In this process, each tree is added to correct the errors of the previous trees. XGBoost is known for its speed, efficiency, and high predictive performance. While building trees, it applies a maximum depth limitation and, when necessary, performs pruning to prevent excessive growth. This helps reduce the risk of overfitting [15].

4.9. Naive Bayes

Naive Bayes is a probabilistic classifier based on Bayes' Theorem, assuming independence between features. According to this algorithm, each feature in the data is independent of the others. It is particularly efficient in problems with a large number of features, enabling fast and effective computation. However, this assumption of independence does not always hold true for real‐world data. Despite this, Naive Bayes performs well in tasks such as text classification and spam detection [15].

4.10. Artificial Neural Networks

Artificial Neural Networks are computational models inspired by the way the human brain processes information. They are composed of layers of nodes, also known as neurons, which are connected to each other in a network‐like structure. Each neuron in a layer receives inputs from the previous layer, processes them, and passes the output to the next layer. Artificial Neural Networks are capable of learning complex patterns in data through a process called backpropagation, where the model adjusts its internal parameters (weights) based on the error between predicted and actual outputs. This iterative process allows the network to improve its predictions over time [16].

4.11. Mobile Application Development

Within the scope of this study, a mobile application named “PC‐BPPV Predictor” was developed to estimate the likelihood of PC‐BPPV. The application was built using the Flutter framework (version 3.10.0), which supports multi‐platform development. As a result, the software can be compiled and run on Windows, Web, Android, and iOS systems.

A Client–Server architecture was adopted to process user input via the application using machine learning algorithms. In this structure, the client (mobile application) communicates with the server over the internet. The user data entered through the app is transmitted to a Web server, which processes the information and returns the results. The server was developed using Python and the Flask framework, with data transferred via the HTTP protocol using the POST method.

Upon launching the application, the main screen displays the app name. Users are prompted to enter information such as age, symptom onset time, symptom duration, trigger factors, and type of dizziness. Based on the submitted data, the application presents one of two outcomes: “PC‐BPPV Positive” or “PC‐BPPV Negative”. A warning message accompanies the result, stating: “This software is based on artificial intelligence (AI) and is part of a pilot study. For a definitive diagnosis and differential diagnosis, you should be examined by a qualified healthcare professional.” A schematic representation of the application's workflow is shown in Figure 1. Performance and verification tests of the mobile application were conducted on Android devices and web‐based platforms using Microsoft Edge and Google Chrome browsers. Android Studio (version: Ladybug 2024.2.1 Patch 2) was used for Android testing. Screenshots of the application showing both PC‐BPPV Positive and PC‐BPPV Negative results are presented in Figure 2A and Figure 2B, respectively. These images were captured using an Android emulator.

FIGURE 1.

Mobile application flow diagram.

FIGURE 2.

A: Mobile application screen for PC‐BPPV Negative result. B: Mobile application screen for PC‐BPPV Positive result.

5. Results

A total of 292 patient records were reviewed. Of the patients examined, 64 (21.9%) had PC‐BPPV, 10 (3.4%) had lateral canal (LC)‐BPPV and 2 (0.7%) had anterior canal (AC)‐BPPV; 28 (9.5%) had persistent postural perceptional dizziness, 30 (10.2%) had probable Meniere's disease, 9 (3.0%) had vestibular neuritis, 44 (15.1%) had orthostatic hypotension, and 1 (0.3%) had multiple sclerosis. One hundred four patients were evaluated as idiopathic because a specific diagnosis could not be reached. Patients diagnosed with LC‐BPPV and AC‐BPPV were not included in the study because the number of data was not sufficient for machine learning analysis. Consequently, data from 280 patients were included in the study. According to these data, patients were divided into PC‐BPPV (n = 64) and other disorders (n = 216).

There was no difference between the PC‐BPPV group and the other disorders group in terms of gender, diabetes mellitus, hypertension, cardiovascular diseases, thyroid, and migraine (p > 0.05). Compared to individuals with other disorders, PC‐BPPV patients were older, had a shorter disease duration, experienced symptoms for a shorter time, had attacks more frequently triggered by specific positional changes, experienced attacks more often in the form of vertigo, and exhibited fewer auditory symptoms (p < 0.05). The characteristics compared according to the groups are presented in Table 2. Therefore, age, dizziness type, triggering factor, onset time, duration of symptoms, and auditory symptom status were selected as features while creating machine learning models.

TABLE 2.

Comparison of demographic and dizziness characteristics of individuals between groups.

Features	PC‐BPPV group, n = 64	Other Disorders, n = 216	Adjusted p‐value
Age (years), median (min‐max)	60 (27–85)	52 (12–83)	0.002a
Gender, n			0.899b
Female	40 (62.6%)	138 (63.9%)
Male	24 (37.5%)	78 (36.1%)
Symptom duration (month), median (min‐max)	0.85 (0.1–12)	1 (0.1–240)	0.002a
Duration of episode, n			0.002b
1–9 s	15 (23.4%)	99 (45.8%)
10–59 s	41 (64.1%)	23 (10.6%)
1–9 min	4 (6.2%)	27 (12.5%)
10–29 min	3 (4.6%)	12 (5.5%)
0.5–24 h	1 (1.5%)	45 (20.8%)
Days	0 (0%)	10 (4.6%)
Dizziness Type, n			0.002b
Vertigo	61 (95.3%)	69 (31.9%)
Presyncope	1 (1.5%)	35 (16.2%)
Lightheadedness	2 (3.1%)	60 (27.8%)
Disequilibrium	0 (0%)	52 (24.1%)
Triggering Factors, n			0.002b
Specific Position	60 (93.7%)	54 (25%)
Sudden Movements	1 (1.5%)	52 (24.1%)
Spontaneous	2 (3.1%)	102 (47.2%)
Continuous	1 (1.5%)	8 (3.7%)
Specific Positions
To stand up, n	42 (65.6%)	142 (65.7%)	0.986b
To go to bed, n	61 (95.3%)	54 (25%)	0.002b
To turn in bed, n	62 (96.8%)	52 (24.1%)	0.002b
Auditory symptom, n	4 (6.2%)	41 (18.9%)	0.028b
Diabetes mellitus, n	16 (25%)	42 (19.4%)	0.515b
Hypertension, n	24 (37.5%)	76 (35.1%)	0.847b
Migraine, n	4 (6.2%)	9 (4.1%)	0.515c
Thyroid, n	0 (0%)	2 (0.09%)	0.810c
Cardiovascular Diseases, n	2 (3.1%)	7 (3.2%)	0.826c

Note: a: Mann–Whitney U test, b: Chi Square, c: Fisher Exact test.

The heat map of the correlation matrix between the features used and the diagnosis is presented in Figure 3A. Among the models created, Random Forest had the highest accuracy value, with 96.43% (Table 3). Therefore, a mobile application was developed based on a Random Forest algorithm. The confusion matrix of the Random Forest algorithm is presented in Figure 3B, and the ROC curve is presented in Figure 3C. Among the inputs used in the Random Forest algorithm, “lying in bed” and “turning in bed” had the highest effect weight. The success of the other models was between 89.28% and 94.64%. Precision, F1‐score, accuracy, recall and ROC‐AUC of all models are presented in Table 3.

FIGURE 3.

A: Correlation coefficients of the features with the result. B: Confusion matrix of the Random Forest model in the testing process. C: ROC curve of the Random Forest model. D: The effect weights of the inputs on the result in the Random Forest algorithm.

TABLE 3.

Test success of the models.

Algorithms	Precision (%)	Recall (%)	F1‐score (%)	Accuracy (%)	AUC‐ ROC (%)
Logistic Regression	80.00	92.31	85.71	92.85	98.74
K‐Nearest Neighbor	81.82	69.23	75.00	89.28	93.47
Support Vector Machine	80.00	92.31	85.71	92.85	98.74
Naive Bayes	76.92	76.92	76.92	89.28	95.88
Decision Tree	75.00	92.31	82.86	91.07	95.08
Random Forest	92.31	92.31	92.31	96.43	98.74
XGBoost	91.67	84.62	88.00	94.64	93.82
Artificial Neural Networks	80.00	92.31	85.71	92.86	93.82

Abbreviation: AUC‐ROC: Area Under the Receiver Operating Characteristic Curve.

6. Discussion

This study aimed to predict PC‐BPPV using vertigo/dizziness features and medical history findings in machine learning. There were differences between PC‐BPPV patients and other disorders patients in terms of age, time of onset of symptoms, duration of symptoms, type of dizziness, accompanying symptoms, and triggering factors. These findings were used in machine learning, and the Random Forest model predicted PC‐BPPV with a 96.4% accuracy rate. In addition, a Random Forest‐based mobile application was developed to predict PC‐BPPV with medical history in our study.

Previous studies indicate that around one quarter of patients who present to hospitals with dizziness symptoms receive a diagnosis of BPPV [17, 18]. Consistent with the literature, 76 (26.0%) of 292 patients in our study had BPPV. AC and LC BPPV symptoms can last longer and be more complex than classic PC‐BPPV symptoms [19]. In addition, these types are less common than PC‐BPPV. Since the number of patients with AC and LC BPPV in our study was insufficient, we did not include them in our study and only included PC‐BPPV patients.

Approximately three‐quarters of dizziness patients can be diagnosed with a medical history [20]. According to Labuguen [21] the type of dizziness should first be determined in patients with dizziness. The type of dizziness can be defined simply by asking the following question; ‘When do you have dizzy spells? Do you feel light‐headed or see the world spin around you?’ Thus, true vertigo and other types of dizziness (presyncope, lightheadedness and disequilibrium) can be distinguished according to the patient's answer. The majority of BPPV patients report true vertigo symptoms [22]. Çakmak et al. [22] compared the symptoms of patients presenting to the emergency department with dizziness complaints with BPPV and other vestibular disorders. The authors stated that 84.2% of BPPV patients complained of vertigo and that vestibular disease is likely to be BPPV in patients presenting to the emergency department with vertigo complaints. Imai et al. [23] stated that dizziness type, attack duration, presence of auditory symptoms, and provoking factors could be used in the detection of BPPV. Consistent with the literatüre [20, 21, 22, 23, 24] we found differences between PC‐BPPV patients and patients with other vestibular disorders in terms of dizziness type, age, attack duration, symptom duration, presence of auditory symptoms, and provoking factors.

The importance of telemedicine, defined as the remote medical support patients receive from physicians, is increasing daily. The COVID‐19 pandemic, in particular, has revealed the importance of telemedicine. BPPV, due to its high prevalence, imposes a significant burden on healthcare systems [25, 26]. Patients with BPPV spend an average of 2685 US dollars per person in the USA [27] and 381 US dollars in Spain [28]. Integrating BPPV into telemedicine can increase the efficiency of healthcare systems and save time and money. Some recent studies have been conducted on this topic [29, 30]. Kim et al. [29] developed a 6‐question questionnaire for self‐diagnosis of BPPV. The authors reported that the questionnaire could detect BPPV with 82.2% accuracy. Another study [31] investigated the feasibility of web‐based diagnosis and treatment of BPPV using a questionnaire and instructional video clips (Video 1). The authors reported a 72.4% accuracy rate for the system. The study emphasized that BPPV is a vestibular disorder suitable for telemedicine and digital treatment. To the best of our knowledge, our study is the first to predict PC‐BPPV using vertigo characteristics and medical history through machine learning, while also developing an Android mobile application based on the most effective algorithm. The present study utilized patients' vertigo/dizziness characteristics and medical history across eight different machine learning models, predicting PC‐BPPV with a 96.4% accuracy rate using the Random Forest model. The questions with the highest impact on the Random Forest algorithm were, respectively, ‘Do you get dizzy when you lie down in bed?’ and ‘Do you get dizzy when you turn over in bed?’ Overall, the accuracy rate of the models ranged from 89.3% to 96.4%. This result demonstrates the high predictability of PC‐BPPV through machine learning using vertigo/dizziness features and medical history.

VIDEO 1.

Mobilemobile application developed to predict PC‐BPPV based on machine learning. Video content can be viewed at https://onlinelibrary.wiley.com/doi/10.1002/lio2.70177

As a tree‐based algorithm, Random Forest can effectively capture the complexity of data by modeling nonlinear relationships and interactions between variables. Furthermore, it is robust in handling common challenges in clinical studies, such as highly correlated predictors and missing data [32]. In the present study, some of the variables used to predict PC‐BPPV, such as vertigo characteristics and medical history, are interactive rather than linear (e.g., “lying down” and “turning in bed”). Random Forest treats these features as separate branches and analyzes the interactions between them. Therefore, the fact that Random Forest emerged as the most successful algorithm in our study may be due to its ability to process nonlinear and interactive data.

The Random Forest model not only achieved the highest performance in accuracy and AUC ROC scores but also excelled across other key performance metrics. Precision reflects the model's ability to correctly identify true positive cases while avoiding false positives, which is critical for reducing unnecessary interventions in clinical practice. Recall measures the model's sensitivity to actual cases, minimizing the risk of overlooking patients with PC‐BPPV. Minimizing false negatives is vital in clinical settings, as undiagnosed BPPV can result in prolonged symptoms, delayed recovery, anxiety, a fear of falling, and residual dizziness [33]. The F1 score provides a balance between precision and recall, offering a comprehensive view of the model's reliability. The combination of strong results across these metrics suggests that the model may be suitable for supporting diagnostic decision‐making in relevant healthcare settings.

A few studies in the literature have used nystagmus features to predict BPPV and its subtypes with machine learning [11, 12]. Wu et al. [11] applied the features of nystagmus observed during the provocative maneuver to 1D and deep learning models in a combined manner to predict BPPV. The study reported that the hybrid algorithm most accurately predicted right and left PC‐BPPV, with AUC‐ROC values of 99%9 and 98%, respectively. Lu et al. [12] used patients' eye movement videos in a multimodal deep‐learning model. The authors reported that the algorithm achieved 81.7% success in diagnosing BPPV. In contrast to previous studies, the present study utilized vertigo‐related characteristics and basic medical history data, which can be obtained without clinical expertise, as input features in machine learning models. Additionally, an Android mobile application was developed based on the best‐performing algorithm. This innovation demonstrates that it may be possible to estimate the likelihood of PC‐BPPV using simple, self‐reported information, making it potentially suitable for use in telemedicine settings. The developed application could enable individuals experiencing vertigo to perform a preliminary assessment via their smartphones before consulting a healthcare provider. This approach may support early awareness and guidance, especially in areas with limited access to healthcare services, and highlights the potential of machine learning to assist in patient‐centered management of vestibular disorders. Future studies could integrate explanatory video clips, including the application of repositioning maneuvers, into the software, enabling research on both the diagnosis and treatment of BPPV using machine learning.

7. Limitations of the Study

This study has several limitations. First, only patient data related to PC‐BPPV were included in the analysis. Data from patients with LC and AC BPPV were not analyzed owing to insufficient sample sizes. Consequently, the high accuracy of the Random Forest model pertains exclusively to PC‐BPPV and limits the generalizability of the findings to other BPPV subtypes. Future research could focus on developing and validating machine learning models that incorporate larger sample sizes and include AC and LC BPPV cases as well. Second, although the developed multi‐platform software is compatible with a variety of systems, it was retrospectively tested using data from a single clinical center. This constraint restricts the external validity of the study. Accurately obtaining a medical history from patients with dizziness or vertigo can be challenging, as symptom‐related terminology often overlaps and may lack appropriate equivalents in certain languages or regions. Notably, the clinician's experience plays a crucial role in interpreting clinical features and patient history, though it may also introduce potential biases during the evaluation process. Therefore, the mobile application should be prospectively (with patients evaluated face‐to‐face) tested by clinicians with varying levels of experience across different clinics and on patients with diverse clinical characteristics. Additionally, to assess the application's feasibility in telemedicine contexts, its usability and reliability in self‐administration by patients should also be examined. Finally, due to the retrospective nature of the study, only a limited set of features could be used to train the machine learning algorithms. Variables such as nausea severity, level of disability, and dizziness/vertigo intensity, which are typically assessed using standardized clinical scales or visual analog scales, were not included. Incorporating these variables in future research may further improve the predictive performance of machine learning models.

8. Conclusion

Our findings demonstrate that dizziness/vertigo characteristics and medical history can effectively be utilized in machine learning to predict PC‐BPPV with high accuracy. Using the Random Forest algorithm, PC‐BPPV was predicted with an accuracy of 96.4%. The mobile application we developed serves as a practical tool for PC‐BPPV prediction in telemedicine. With its multi‐platform compatibility, the software can operate on devices with any operating system, enhancing accessibility. The high accuracy rate of the mobile application in predicting PC‐BPPV highlights the potential for AI platforms to contribute to vestibular science within telemedicine. This innovation may also help raise awareness about the use of AI in this field.

Ethics Statement

Ethical approval for this retrospective study was obtained from the Karabük University Non‐Interventional Ethics Committee (Decision No: 2024/1650). Written approval was received from all subjects involved in the study.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.