This study was a retrospective case analysis and included 522 patients with vertigo diagnosed in the SRM-IV vertigo diagnostic and treatment system. Patient data were derived from the system’s big data platform and spanned a 1-year period (January 2024–January 2025). The study was approved by the Ethics Committee of the Affiliated Hospital of Shandong Second Medical University. The inclusion criteria were (1) meeting the diagnostic criteria for vertigo in the Vertigo Diagnostic and Treatment Guidelines, and (2) having a complete clinical history, including demographic information (gender, age, marital status, occupation, and level of education), medical history (duration of the disease, side of the onset of the disease, comorbidities, etc.), laboratory indices (white blood cell count, neutrophil count, lymphocyte count, monocyte count, C-reactive protein, ferritin, IL6) and the degree of hearing impairment. The exclusion criteria were (1) incomplete clinical data or missing key information, and (2) the presence of other definite etiological factors that could lead to vertigo, such as central nervous system diseases and drug intoxication. Based on the diagnosis, patients were divided into BPPV group (n = 303) and non-BPPV group (n = 219). All data were desensitized to protect patient privacy.

To identify the key feature variables for BPPV diagnosis, the following machine learning methods were used in this study:(1) LASSO regression analysis: The LASSO regression model was applied for feature selection. The optimal penalty coefficient λ is determined through cross-validation and the corresponding feature variable at Lambda.1se is selected. (2) Random Forest: The randomForest package was used, setting ntree = 1,000 and mtry = 3 (square root of the number of predictor variables). The importance of the variables was assessed using both Gini coefficient decline (type1) and replacement importance (type2) methods, and the top 10 ranked variables were selected. To avoid overfitting, out-of-bag (OOB) errors were used to assess model performance.

Multifactor Logistic regression analysis was performed based on the feature variables screened by machine learning. Whether the diagnosis of BPPV was made was used as the dependent variable, and the above feature variables were included in the model as independent variables. Statistically significant variables were screened using stepwise regression and the respective odds ratios (OR) and 95% confidence intervals (CI) were calculated. Based on the results of the multifactorial logistic regression analysis, a predictive model for BPPV-assisted diagnosis was constructed. The model was presented in the form of a column-line graph for easy clinical application.

The established prediction model was applied to another batch of BPPV patients diagnosed in the SRM-IV vertigo diagnosis and treatment system (validation set), and the prediction effect of the model on an independent dataset was evaluated. The validation set was an independent cohort (n = 112) of patients from subsequent visits to the same center. (1) Receiver Operating Characteristic (ROC) Curve and AUC: ROC curves were plotted and AUC was calculated to assess the discriminatory power of the model. The higher the AUC value, the higher the prediction accuracy of the model. (2) Clinical calibration curves: The calibration of the model is assessed by plotting clinical calibration curves comparing the difference between the predicted probability and the actual incidence. The closer the calibration curve is to the diagonal, the better the model’s predictions match the actual situation. (3) Clinical Decision Curve and Clinical Impact Curve: Draw the clinical decision curve and clinical impact curve to evaluate the clinical net benefit rate and patient benefit of the model under different threshold probabilities to evaluate the clinical application value of the model.

SPSS 26.0 and R 4.1.0 software were used for statistical analysis. Normally distributed continuous data were expressed as mean ± standard deviation (Mean ± SD), and T-test was used for comparison between groups. Non-normally distributed continuous data were presented as median (interquartile range, Median [Q1, Q3]) and Mann–Whitney U test was used for comparison between groups. Numerological data were expressed as the number of cases [percentage, n (%)] and comparisons between groups were performed using either the chi-square test or Fisher’s exact test. p < 0.05 was statistically significant.

A total of 522 subjects were included in this study, of which 303 were diagnosed with BPPV and 219 were non-BPPV. There was no significant difference in gender composition between the two groups (p = 0.9137; Table 1), but there was a significant difference in age distribution (p < 0.0001), and the mean age of the BPPV group was higher than that of the non-BPPV group (57.7 ± 19.0 vs. 50.8 ± 14.6). There was also a significant difference between the two groups in terms of disease duration (p < 0.0001), with a median disease duration of 4.88 days [0.75; 9.33] in the BPPV group and 7.87 days [0.49; 14.8] in the non-BPPV group.

In terms of laboratory indicators, there was no significant difference in white blood cell count between the two groups (p = 0.1948), but there were significant differences in neutrophil (p = 0.0095), lymphocyte (p = 0.0182), monocyte (p = 0.0001), CRP (p < 0.0001), ferritin (p = 0.0023), and IL-6 (p < 0.0001). The levels of monocytes in the BPPV group were higher than those in the non-BPPV group, while the levels of CRP, ferritin, and IL-6 were lower than those in the non-BPPV group (Table 2).

There was no significant difference in the degree of hearing loss between the two groups (p = 0.4563). There were also no significant differences in incidence (p = 0.3679), marital status (p = 0.2836), occupation (p = 0.8054), and education level (p = 0.1031) between the two groups. In terms of comorbidities, anxiety (p < 0.0001) and depression (p < 0.0001) were significantly different between the two groups, and the incidence of anxiety and depression was higher in the non-BPPV group than in the BPPV group. Vitamin D deficiency was more common in the non-BPPV group (p < 0.0001). Stroke (p = 0.0242), diabetes mellitus (p = 0.0386), hyperuricemia (p = 0.0242), and history of head surgery (p = 0.0234) were significantly different between the two groups, and were more common in the non-BPPV group. There were no significant differences in essential hypertension (p = 0.0687), vestibular migraine (p = 0.0836), head trauma (p = 0.2150), insomnia (p = 0.3590), smoking (p = 0.0900), and alcohol consumption (p = 0.6556) between the two groups. Otitis media was more common in the non-BPPV group (p = 0.0001; Table 2).

Based on the LASSO regression analysis, a total of 9 core targets were obtained at Lambda$1se, and random forests were used for variable screening using type1 and type2 to obtain the top 10 variables in the sorting, respectively, (Figure 1). The Venn diagram ultimately acquires eight core targets for machine learning: disease course, leukocytes, neutrophils, lymphocytes, monocytes, CRP, ferritin, and vitamin D deficiency (Figure 2).

Based on machine learning, disease duration, white blood cell, neutrophils, lymphocytes, monocytes, CRP, ferritin, and vitamin D deficiency factors were finally obtained and included in the multifactorial logistic regression analysis (Table 3). The results showed that disease duration, leukocytes, neutrophils, lymphocytes, mono-CRP, ferritin, and vitamin D deficiency, were independent risk factors for the diagnosis of BPPV (OR>1, p < 0.05), and monocyte count was an independent protective factor for the diagnosis of BPPV (OR<1, p < 0.05). Based on multifactor logistic regression analysis, a prediction model was established as shown in Figures 3, 4.

The established prediction model was externally applied to the patients with BPPV diagnosed by the SRM-IV vertigo diagnosis and treatment system, and the ROC curve was plotted and the AUC was calculated to be 0.927. Further plotting of clinical calibration curves showed that the predicted values were in good agreement with the actual values. Finally, the clinical decision curve and clinical impact curve were plotted, and the established prediction model had a good clinical net benefit rate (Figure 5).

In this study, 522 patients with vertigo were included, of which 303 were diagnosed with BPPV and 219 with non-BPPV. Among the 303 BPPV patients, about 76% were posterior canals, 21% were horizontal canals, and 3% were superior canals. The rate of nystagmus disappearance after reduction was 92.7% (281/303). There was no significant difference between the two groups in terms of gender composition, but there was a significant difference in age distribution and duration of the disease. The mean age of the patients in the BPPV group was higher than that of the non-BPPV group. It has been shown that the incidence of BPPV increases with age, which may be related to the deterioration of vestibular function in the elderly (12). In addition, this study found that patients in the BPPV group had a shorter duration of illness than those in the non-BPPV group, suggesting that BPPV may be characterized by a more acute onset.

In terms of laboratory indices, there were significant differences between the two groups in several indices. Patients in the BPPV group had higher levels of monocytes than those in the non-BPPV group, whereas CRP, ferritin and IL-6 levels were lower. These results suggest that inflammatory response and iron metabolism disorders may play an important role in the pathogenesis of BPPV. It has been found that inflammation of the vestibular system can lead to otolith dislodgement and trigger vertigo symptoms (13). Iron metabolism disorders have also been reported to be associated with vestibular dysfunction (14).

In terms of comorbidities, patients in the non-BPPV group had a higher incidence of anxiety and depression than the BPPV group, and vitamin D deficiency was also more common. This suggests that psychological factors and vitamin D levels may be important clues in the differential diagnosis of BPPV (15, 16). It has been shown that mood disorders such as anxiety and depression can cause vestibular dysfunction leading to vertigo symptoms (17, 18). Vitamin D deficiency has also been found to be associated with impaired vestibular function (19). In addition, stroke, diabetes mellitus, hyperuricaemia and history of head surgery were more common in the non-BPPV group, suggesting that these factors may be potential causes of non-BPPV vertigo.

In this study, two machine learning methods, LASSO regression and random forest, were used to screen out feature variables that are valuable for BPPV diagnosis. LASSO regression obtained 9 core targets at Lambda$1se, while random forest obtained the top 10 ranked variables, respectively. Integration of the results of the two methods through Venn diagrams resulted in the identification of eight core features, including disease duration, white blood cell count, neutrophil count, lymphocyte count, monocyte count, CRP, ferritin, and vitamin D deficiency.

These results are consistent with the findings of the aforementioned baseline data analysis, further confirming the important roles of inflammatory response, iron metabolism disorders and vitamin D deficiency in the pathogenesis of BPPV. Meanwhile, machine learning methods have also revealed the value of clinical indicators such as disease duration and blood cell counts in the diagnosis of BPPV. One study used decision trees and random forests to construct a diagnostic model for vertigo diseases, and found that factors such as age, gender, and the nature of vertigo contributed most to the discriminative power of the model (5). Another study applied support vector machines and artificial neural networks to analyze the results of vestibular function tests to achieve automatic classification and diagnosis of vertigo diseases (20). These studies show that machine learning is expected to uncover key information for diagnosing vertigo diseases from massive clinical data and provide powerful support for clinical decision-making.

On the basis of the eight core features screened by machine learning, this study further used multifactorial logistic regression analysis to construct an auxiliary diagnostic model for BPPV. The results of regression analysis showed that disease duration, white blood cell count, neutrophil count, lymphocyte count, CRP, ferritin and vitamin D deficiency were independent risk factors for the diagnosis of BPPV, while monocyte count was an independent protective factor. These findings are consistent with the previous discussion, suggesting that inflammatory responses, disturbances in iron metabolism and vitamin D deficiency may be involved in the pathogenesis of BPPV by affecting vestibular system function.

Based on the logistic regression model, this study drew column-line graphs to visualize the relationship between factors and the risk of BPPV, which provides a convenient tool for clinicians to estimate the probability of developing BPPV based on the patient’s specific situation. One study used multifactorial logistic regression to construct a prediction model for BPPV and found that age, duration of vertigo, vertigo triggers and vestibular function test results were independent predictors of BPPV (21). This is partially similar to the findings of this study, suggesting that clinical characteristics and vestibular function indicators are valuable for the diagnosis and prediction of vertigo disorders.

In order to assess the actual performance of the constructed auxiliary diagnostic model, external validation was performed in this study using data from patients with BPPV diagnosed by the SRM-IV Vertigo Diagnostic and Treatment System. ROC curve analysis showed that the model had a high predictive accuracy for BPPV, with an AUC of 0.927. Calibration curve analysis also confirmed the good agreement between the model’s predicted probability and the actual incidence rate. In addition, decision curve and clinical impact curve analyses showed that the model could provide a net benefit for clinical decision making and had good clinical application value.

These results support the usefulness of the auxiliary diagnostic model constructed in this study in the diagnosis of BPPV. By incorporating multidimensional information such as clinical features and laboratory indicators, the model is able to provide an objective and quantitative basis for the diagnosis of BPPV, which helps to improve the accuracy and efficiency of diagnosis. Similarly, one study used machine learning to construct a prediction model for BPPV and validated it on an external dataset, confirming the reliability and usefulness of the model (22, 23). Another study integrated a machine-learning diagnostic model for vertigo disease into a clinical decision support system and achieved good application results (24). These studies show that the auxiliary diagnostic models are expected to provide intelligent tools for the diagnosis and treatment of vertigo diseases, reduce the workload of doctors, and improve the level of diagnosis and treatment.

However, this study was a retrospective single-center design with selection bias. Potential variables such as serum calcium/bone metabolism indexes were not included; there were differences in the time intervals between laboratory tests and symptom onset; and the model did not incorporate data on quantitative assessment of vestibular function and response to treatment, which limited its value for clinical application. Based on the findings of this study, future research should conduct a multicenter prospective study to validate the external validity of the model; in addition, the relationship between the dynamic changes in biomarkers and the prognosis of BPPV should be explored; and clinical, laboratory, and imaging data should be integrated to construct a multimodal prediction model. In-depth mechanistic studies should be conducted to elucidate the molecular pathways of inflammation and metabolic disorders leading to BPPV.