Machine learning models for predicting transstenotic pressure gradient based on computed tomography angiography features

Introduction

The transstenotic pressure gradient (TPG) is an important indicator of the severity of cerebral venous reflux obstruction (1). TPG caused by transverse sinus stenosis (TSS) plays a critical role in conditions such as intracranial hypertension (IIH) and pulsatile tinnitus (PT) (2,3). Venous sinus stenting is an effective treatment for IIH and PT, as it can significantly reduce intracranial pressure (4,5). A TPG value from the superior sagittal sinus to the jugular bulb exceeding 8 mmHg is often used as a criterion for stenting (6), while a value of approximately 4 mmHg is typically considered a normal pressure gradient and found in healthy individuals (7). Currently, digital subtraction angiography (DSA) is the gold standard for TPG measurement; however, its invasiveness limits its widespread use. Therefore, developing noninvasive methods to predict TPG is essential.

Computed tomography angiography (CTA), a high-resolution imaging modality, is widely used to evaluate transverse sinuses. Using CTA data, Zhao et al. (8) demonstrated that the extent and degree of TSS on the affected side, as well as contralateral transverse sinus dysplasia, were independently correlated with the relative pressure inside the sinus cavity. A multivariate pressure prediction model based on TSS morphology was subsequently proposed. However, traditional methods face limitations in fully capturing the rich information contained in CTA images. Machine learning is increasingly being applied in the predictive modeling of cardiovascular conditions. For instance, Sunderland et al. (9) combined computational fluid dynamics with support vector machines (SVMs) to identify blood-flow vortex intensity as an independent risk factor for intracranial aneurysm rupture. Similarly, Cai et al. (10) developed machine learning models to predict cerebral blood flow in the internal carotid artery.

Therefore, we aimed to construct a machine learning model based on relevant CTA indicators from patients who had undergone DSA and TPG measurement capable of noninvasively predicting the pressure gradient across the TSS. We anticipate that our findings will support individualized treatment planning and guide clinical decision-making. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-157/rc).

Methods

Participants

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and was approved by the Ethics Committee of Beijing Friendship Hospital, Capital Medical University (approval Nos. 2020-P2-202-02 and 2023-P2-095-01). The requirement for individual informed consent was waived due to the retrospective nature of the analysis.

We retrospectively included patients who underwent DSA at Beijing Friendship Hospital between January 2017 and December 2023. The inclusion criteria were (I) completion of CTA and (II) a diagnosis of TSS, defined as a restrictive or segmental filling defect in the transverse sinus lumen. The degree of TSS was calculated as the ratio of the stenotic diameter to the normal lumen diameter. Only patients with >50% TSS (11) were included. Meanwhile, the exclusion criteria were as follows: (I) imaging evidence of aneurysms, arteriovenous malformations/fistulas, abnormal vascular communications, or tumors; (II) inflammatory ear lesions; (III) other diseases affecting intracranial hemodynamics; and (IV) incomplete clinical or laboratory data. To meet the sample size requirement for building machine learning models (at least 10 times the number of features), we ensured adequate patient numbers. Patients who underwent DSA at Beijing Tiantan Hospital and met the above criteria were included for external validation of the machine learning model.

CTA examination

CTA scans were performed with either a 64-slice spiral CT device (Brilliance, Philips Healthcare, Best, The Netherlands) or a 256-slice spiral CT device (Revolution, GE HealthCare, Chicago, IL, USA), covering the range from the sixth cervical vertebra to the cranial vertex. The scanning parameters are listed in Table 1. The threshold for trigger bolus tracking was 150 Hounsfield units (HU). The contrast agent iopamidol (370 mg/mL) was administered through the anterior cubital vein at a dose of 1.5 mL/kg and an injection rate of 5 mL/s. This was followed by an injection of 20 mL of physiological saline at the same rate. Scanning commenced 35–40 s after contrast administration, extending from the cranial vertex to the skull base.

CTA image analysis

CTA data in Digital Imaging and Communications in Medicine (DICOM) format were transferred to the postprocessing workstation for curved planar reformation. The measurement indicators and definitions are described below. The measurement methods for each indicator were based on previous studies (8,12,13). Data were measured collaboratively by two senior radiologists.

Residual area ratio

The cross-sectional area of the residual lumen at the TSS on the symptomatic side was measured on curved planar reformation. The normal area of the transverse sinus—defined as the lumen area at the distal, nonstenotic segment, with the confluence of the cortical veins being excluded—was also measured. The residual area ratio was calculated as follows = residual area ratio = cross-sectional area of the residual lumen/normal cross-sectional area (details of the measurements are provided in Figure 1A-1C).

Figure 1 Measurements of four features based on computed tomography angiography. The symptomatic transverse sinus is shown in a curved reformatting CT image. A reformatting image perpendicular to the stenosis (A; section 1) was obtained, and the area of the lumen at the stenosis was measured (B). The normal TS area was obtained in the same manner (A; section 2, and C). The residual area ratio of this patient was 0.45 (14 mm²/31 mm²) at this site. The stenosis length was measured at 8.1 mm (A; white dotted line). Intrinsic stenosis is defined as a localized filling defect in the transverse sinus (D; black arrow), while extrinsic stenosis appears as a long and smooth cone-shaped narrowing (E; black arrow). The Labbé vein is shown joining the proximal end of the TSS (A; black arrow), the stenosis segment (D; white arrow), and the distal end (E; white arrow). CT, computed tomography; TS, transverse sinus; TSS, transverse sinus stenosis.

Stenosis length

The length of the stenosis on the symptomatic side was measured via curved planar reformations. In cases with multiple stenoses, the most severe one was selected (details of the measurements are provided in Figure 1A).

Stenosis type

Based on morphology, TSS was classified as (I) intrinsic stenosis (localized filling defects in the transverse sinus, often caused by arachnoid granule compression); (II) extrinsic stenosis (manifesting as a long and smooth cone-shaped stenosis, often caused by compression of external veins or brain parenchyma); and (III) mixed stenosis (combination of intrinsic and extrinsic forms). Examples of these types are presented in Figure 1D,1E.

Location of the Labbé vein

The confluence location of the Labbé vein relative to the TSS was classified as follows: 1= lower (proximal to the stenosis), 2= middle (at the stenosis), and 3= upper (distal to the stenosis). Examples of these types are presented in Figure 1A,1D,1E.

Degree of drainage dominance

The cross-sectional area of the transverse sinus on the symptomatic and contralateral sides was measured, and the degree of drainage dominance was calculated as follows: degree of drainage dominance = cross-sectional area of the symptomatic transverse sinus/(cross-sectional area of the symptomatic transverse sinus + cross-sectional area of the contralateral transverse sinus). The details of the measurement method are provided in Figure 2A-2C.

Figure 2 Measurements of two additional features based on computed tomography angiography. CT images of a 44-year-old female with right-sided pulsatile tinnitus are shown. The bilateral transverse sinuses are displayed in a curved reformatting CT image (A). Stenosis was also present in the transverse sinus on the nonsymptomatic side (A; black arrow). Image reformation was performed perpendicular to the symptomatic (A; section 1) and contralateral (A; section 2) TS without stenosis. The normal areas of the symptomatic (B) and contralateral (C) TS were measured. The degree of drainage dominance was 0.72 [42 mm²/(42 mm² + 16 mm²)]. The stenosis area of the contralateral TS was measured in similar fashion (A; section 3, and D), with a degree of contralateral stenosis of 0.75 (1–4 mm²/16 mm²). CT, computed tomography; TS, transverse sinus.

Degree of contralateral stenosis

If stenosis was present on the contralateral side, its residual ratio was calculated. The degree of contralateral stenosis was calculated as 1—the residual area ratio of the contralateral stenosis. If no stenosis was present, the value was recorded as 0. The details of the measurements are provided in Figure 2A,2C,2D.

DSA examination

Under local anesthesia, DSA was performed with the Seldinger technique and an angiography system (Innova 4100 IQ, GE HealthCare) within 7 days of CTA. The patient was supine, and a percutaneous puncture of the right femoral artery was performed. Subsequently, a 5F catheter was advanced into the left and right internal carotid arteries and the left vertebral artery. A second percutaneous puncture was made in the right femoral vein, through which a 5F catheter was introduced to the symptomatic internal jugular vein. A 2.7F microcatheter was used for superselection and inserted into the intracranial venous sinus, reaching the distal end of the TSS and the proximal end of the sigmoid sinus on the affected side. A pressure sensor connected to the microcatheter measured the TPG.

Machine learning model construction and statistical analysis

TPG was used as the dependent variable, and six CTA-derived features, serving as independent variables, were used to construct binary classification models based on thresholds of 4 and 8 mmHg. Additionally, we attempted to develop a three-class classification model based on these two thresholds. Six machine learning models were implemented with Python version 3.9.0 (Python Software Foundation, Wilmington, DE, USA) and sci-kit-learn library version 1.5.2 to predict TPG in the transverse sinus. The operating environment was Intel Core i7-4500u (Intel, Santa Clara, CA, USA) with a 2.3 GHz processor and 5 GB RAM running memory in Windows 10. The six machine learning algorithms are outlined in the following sections.

Logistic regression

Logistic regression is a linear classification algorithm commonly used in binary and multiclassification tasks. In this study, multiclassification tasks were extended by one-vs-rest, one-vs-one, and multinomial or softmax approaches.

Naïve bayes (NB)

NB is a simple and effective classification method based on probability theory, which assumes that each feature is independent of the others.

k-nearest neighbors (KNN)

KNN is a nonparametric classification algorithm that classifies new samples by comparing their distances to the KNN samples in the training set. Commonly used distance measures include the Euclidean and Manhattan distances.

Support vector machine (SVM)

SVM is a supervised learning algorithm used for classification and regression tasks, which separates different classes of samples by determining decision boundaries with maximum intervals. In cases where the original feature space is not separable, SVM uses different kernel functions (e.g., linear kernel, polynomial kernel, and radial basis function kernel) to map the data to a higher-dimensional space, making it linearly separable.

Adaptive boosting (AdaBoost)

AdaBoost is an enhancement method (boosting) in ensemble learning that forms a strong classifier by combining multiple weak classifiers (e.g., decision trees). During the training period, it pays more attention to the samples that have been incorrectly classified in the previous classification and gradually improves the accuracy of the classification.

Light gradient boosting machine (LightGBM)

LightGBM is an efficient implementation of gradient-boosted decision trees (GBDTs) based on gradient boosting, designed for processing large-scale data and high-dimensional features. Compared with other GBDT frameworks, LightGBM significantly improves training speed and memory efficiency through use of a histogram-based algorithm.

Evaluation of model performance

We conducted fivefold cross-validation on the training dataset to evaluate model performance. Cross-validation is a widely used technique in machine learning because it enables the estimation of models with low bias. To address class imbalance in classification tasks, we implemented an inverse-frequency weighting mechanism during training. The weight for each class equals the total sample count divided by its class frequency. This weighting is applied to the loss function, increasing penalty for misclassifying minority-class samples and forcing the model to focus more on challenging minority cases. The approach effectively mitigates gradient bias from class imbalance while improving sensitivity to minority classes and overall model robustness. Model performance was evaluated according to accuracy, sensitivity, specificity, F1-score, Matthews correlation coefficient (MCC), and area under the curve (AUC), as described in Eqs. [1-6]. Subsequently, we employed Shapley additive explanations (SHAP) analysis to further investigate the principal factors influencing TPG. Finally, we used an external dataset to validate and determine model quality through accuracy.

Accuracy=TP+TNTP+FP+FN+TN×100%[1]

Precision=TPTP+FP×100%[2]

Sensitivity=TPTP+FN×100%[3]

Specificity=TNTN+FP×100%[4]

F1score=2×Precision×RecallPrecision+Recall×100%[5]

MCC=TP×TN−FP×FN(TP+FP)×(TP+FN)×(TN+FP)×(TN+FN)[6]

where TP is the number of true-positive cases, TN the number of true-negative cases, FP the number of false-positive cases, and FN the number of false-negative cases.

We also applied a grid search to optimize hyperparameters for each model. For LR, penalty =12; C =1.0, solver = lbfgs, and max_iter =1,000. For NB, var_smoothing =1e−9. For KNN, _neighbors =5 and ​= distance. For SVM, kernel = rbf and C =1. For AdaBoost, n_estimators =10, learning_rate =0.2, and Max_depth =5. For LightGBM, n_estimators =10, Learning_rate =0.2, num_leaves =32, and Max_depth =−1.

Results

Baseline demographics

A total of 139 patients were included, comprising 18 men and 121 women, with a mean age of 39.16±11.83 years. The mean TPG was 7.25±4.22 mmHg. Based on the 4-mmHg threshold, 35 patients had TPG <4 mmHg, and 104 had TPG >4 mmHg. Based on the 8-mmHg threshold, 84 patients had TPG <8 mmHg, and 55 had TPG >8 mmHg. The general characteristics and measurement indicators of patients are summarized in Table 2.

Machine learning model performance

Using CTA features, we built six machine learning models to classify whether TPG exceeded the 4-mmHg threshold and validated each model’s performance. Table 3 summarizes the experimental results for each model. The LR model achieved the highest accuracy (0.81), followed by the SVM (0.8), KNN (0.8), NB (0.76), LGB (0.75), and AdaBoost (0.75) models, indicating that machine learning based on the CTA prediction model can correctly distinguish the TPG of the transverse sinus. The KNN model showed high sensitivity and a high F1-score (0.7 and 0.8, respectively), while the LR model had the highest MCC (0.45). However, the overall specificity of the six models was low, with LR performing the best (0.46). Receiver operating characteristic (ROC) curves with 95% confidence intervals for the different models were plotted to reflect further the prediction results of each model (Figure 3A). The AUC values for LR, AdaBoost, NB, LGB, SVM, and KNN were 0.83, 0.60, 0.81, 0.74, 0.80, and 0.70, respectively. The LR model showed the largest AUC and most optimal performance among all the prediction models.

Figure 3 Performance of the different models evaluated according to ROC curves (A) based on a 4-mmHg threshold and (B) based on an 8-mmHg threshold. Ada, adaptive boosting; AUC, area under the curve; CV, cross validation; KNN, k-nearest neighbors; LGB, light gradient boosting; LR, logistic regression; NB, naïve bayes; ROC, receiver operating characteristic; SVM, support vector machine.

Similarly, we constructed a binary classification model based on the 8-mmHg threshold using the six CTA features. Table 4 summarizes the experimental results for each model. Among the six models, the LR model exhibited the highest accuracy (0.79). The accuracies of AdaBoost, NB, LGB, SVM, and KNN were 0.66, 0.73, 0.73, 0.78, and 0.71, respectively. The sensitivity and specificity of the six models showed a balanced performance, and the overall value was >0.6, with the NB model having the highest specificity (0.85) and LGB having the highest sensitivity (0.82). We also found that the LR and SVM models had the best MCC (0.56), and the LR model had the highest F1-score (0.75). Subsequently, we compared the diagnostic efficacies of the different models using ROC curves. The larger the AUC was, the better the model classification. The AUCs for LR, AdaBoost, NB, LGB, SVM, and KNN were 0.83, 0.73, 0.83, 0.77, 0.83, and 0.79, respectively (Figure 3B). LR, NB, and SVM showed optimal performance across the six prediction methods, with the largest AUCs (0.83).

In the three-class classification model, AdaBoost achieved the highest accuracy (0.65) and MCC (0.41). The remaining models showed lower performance metrics (see Table S1 for details). Due to the suboptimal overall performance of the three-class model, no further interpretability analysis was conducted.

Model interpretability analysis

SHAP analysis was used to evaluate the impact of each feature on TPG and improve model interpretability. In the 4-mmHg binary classification model, the LR model (with the highest AUC) was analyzed. The top three most important features were residual area ratio, degree of drainage dominance, and location of the Labbé vein (Figure 4A). Among these, the location of the Labbé vein, contralateral stenosis, and stenosis type were positively correlated with TPG, while residual area ratio, degree of drainage, and stenosis length were negatively correlated (Figure 4B).

Figure 4 Interpretability analysis of the binary LR model based on the 4-mmHg threshold. (A) Bar chart of the importance ranking of the features. (B) SHAP feature density scatter plot. The x-axis shows SHAP values. Point colors represent feature values (red = high, blue = low). Positive SHAP values increase the predicted probability of TPG >4 mmHg, while negative values decrease it. LR, logistic regression; SHAP, Shapley additive explanations; TPG, transstenotic pressure gradient.

In the 8-mmHg binary classification model, the ROC was the same as that of the LR, NB, and SVM models. Considering that the LR model had the highest accuracy, we arbitrarily analyzed it for interpretability. In the special importance analysis, the six features in descending order of importance were residual area ratio, degree of drainage dominance, location of the Labbé vein, stenosis length, stenosis type, and degree of contralateral stenosis (Figure 5A). Figure 5B shows that among these characteristics, the location of the Labbé vein, stenosis length, stenosis type, and degree of contralateral stenosis were positively correlated with TPG, while the residual area ratio and degree of drainage dominance were negatively correlated.

Figure 5 Interpretability analysis of the binary LR model based on the 8-mmHg threshold. (A) Bar chart of the importance ranking of the features. (B) SHAP feature density scatter plot. The x-axis shows SHAP values. Point colors represent feature values (red = high, blue = low). Positive SHAP values increase the predicted probability of TPG >8 mmHg, while negative values decrease it. LR, logistic regression; SHAP, Shapley additive explanations; TPG, transstenotic pressure gradient.

External validation

The external dataset comprised 19 patients (4 males and 15 females), with a mean age of 37.68±12.07 years. For the 4-mmHg threshold classification model, the KNN and NB models achieved external validation accuracy up to 0.89 (Figure 6A). For the 8-mmHg model, the AdaBoost model also achieved an external validation accuracy of 0.89 (Figure 6B).

Figure 6 Accuracy of external validation of machine learning models (A) based on a 4-mmHg threshold and (B) based on an 8-mmHg threshold. AdaBoost, adaptive boosting; KNN, k-nearest neighbors; LGB, light gradient boosting; LR, logistic regression; SVM, support vector machine.

Discussion

In this study, six features were extracted from CTA images, with TPG induced by sinus stenosis serving as the dependent variable. Subsequently, six binary machine learning classification models were constructed with thresholds of 4 and 8 mmHg, respectively. These models demonstrated good performance, with the LR model yielding the best results. Furthermore, interpretability analysis identified the residual area ratio, degree of drainage dominance, and location of the Labbé vein as the three most important features in the LR model, with the residual area ratio and degree of drainage dominance being negatively correlated with TPG and the location of the Labbé being positively correlated. Finally, the model’s high accuracy was confirmed through external validation.

Among the six models developed, traditional machine learning algorithms including LR, NB, and SVM outperformed the tree-based models. In particular, LR achieved the best performance across nearly all evaluation metrics. This may suggest a linear relationship between features and the target variable. LR can iteratively identify the strongest linear combination of variables (14). It generates a simple linear equation that relates predictors to the log odds of the outcome, offering high interpretability. Additionally, in imbalanced datasets, complex models may overfit the majority class, whereas LR handles imbalances more effectively through linear discrimination and probabilistic estimation (15,16). In clinical practice, LR models are widely used for predicting complications after duodenectomy (17) and cognitive impairment among intensive care unit inpatients (18). These findings indicate that LR is as effective as are other complex machine learning models. Overall, LR is the most suitable method for analyzing binary classification tasks, as it has high diagnostic power and superior interpretability, which are more important than is predictive performance in clinical scenarios. Finally, the LR model’s compact size and exceptional version stability allow efficient multiplatform deployment, making it well-suited for resource-limited environments and facilitating the broader application of the model developed in this study.

Venous sinus stenting is an effective method for treating IIH and PT. The venous pressure gradient is used to assess stenting suitability. The gold standard for measuring the venous pressure gradient across the TSS is cerebral venography with manometry via DSA, an invasive procedure. However, the optimal TPG threshold for venous stenting remains controversial. Reported thresholds during stent implantation range from a minimum of >4 mmHg to a maximum of >10 mmHg (19,20). In most studies, TPG >8 mmHg has been considered the clinical threshold for stent placement (6,11,21). Additionally, the pressure gradient from the superior sagittal sinus to the jugular bulb in healthy individuals is approximately 4 mmHg (7), and gradients below this are typically not treated. Therefore, we used 4 and 8 mmHg as thresholds to construct the binary classification model: <4 mmHg was considered normal, >4 mmHg indicated a potential need for stenting, and >8 mmHg indicated a strong need for surgical intervention.

Recently, interest has grown in the noninvasive prediction of venous sinus TPG. Zhang et al. developed an SVM model using arteriographic perfusion map-derived radiomic features to predict TPG, achieving accurate predictions at 8 and 6 mmHg (22). Ma et al. developed a radiomics model based on magnetic resonance venography to predict a high TPG in patients with IIH diagnosed with venous sinus stenosis, reporting an AUC of the test dataset as high as 0.877 (23). In contrast, we used features from CTA data to construct six machine learning models to predict TPG, among which the LR model exhibited the best performance, with an AUC of 0.83. We also examined a three-class model using thresholds of 4 and 8 mmHg, but due to further reductions in sample size, the overall accuracy fell below 0.65. As a result, we did not conduct further interpretability analysis. In the future, we aim to develop a semiautomated software module can be developed on the 3D Slicer platform to package trained model files within the module’s code directory and to employ Slicer’s Python environment to invoke the corresponding framework for model loading and inference. The module will support importing head CTA images, leverage Slicer’s visualization capabilities for image display, automatically extract transverse sinus structural features, and provide an interactive interface for physicians to manually supplement additional characteristics. The extracted data will then be fed into the built-in classification model to output real-time TPG prediction results, along with confidence scores, thereby integrating the TPG prediction model into the medical imaging workflow.

The indicators used in this study were selected based on prior clinical experience. To avoid collinearity, we excluded variables with high internal correlations—such as that between the type of drainage dominance and degree of drainage dominance—and ultimately included six CTA features. We identified three important features: residual area ratio, degree of drainage dominance, and location of the Labbé vein. The residual area ratio reflects the severity of stenosis: the smaller the residual area ratio is, the greater the stenosis severity and the larger the TPG. For every 10% increase in venous sinus stenosis in patients with IIH, the pressure gradient increases by approximately 3.5 mmHg (24). The drainage-dominant side represents the side with the highest blood flow, and the degree of drainage dominance reflects blood flow on the symptomatic side. The larger the degree of drainage dominance is, the higher the blood flow on the symptomatic side. PT often occurs on the dominant side, suggesting that venous PT may be related to high flow (25). This may be because high blood flow is more likely to generate larger TPGs through stenosis, leading to hemodynamic changes at the stenosis site and downstream. The Labbé vein is a component of the superficial cerebral vein, originating from the Sylvian fissure and flowing backward into the transverse sinus, serving as an anastomotic vein between the superficial middle cerebral vein and the transverse sinus (26). We found that the closer the Labbé vein confluence point was to the distal end, the larger the TPG. Similarly, Zhao et al. (8) found that when the Labbé vein merged with the proximal end of the TSS, the TPG was relatively small.

Limitations

Despite promising results, this study involved certain limitations which should be addressed. First, the analysis included only six CTA indicators, and other relevant features might have been excluded. Second, the overall sample size and group sizes within the binary classifications were relatively small and unbalanced. Finally, the external validation dataset was limited in size. Therefore, multicenter studies with a large sample should be conducted in the future to confirm the model’s generalizability and robustness.

Conclusions

In this study, a machine learning model based on CTA features was successfully developed for predicting transverse sinus TPG. The model requires only six CTA features and is, therefore, easy to implement. The results demonstrated the feasibility of predicting the transverse sinus TPG through use of such CTA data. This approach may help clinicians to determine whether patients with IIH or PT require DSA examination, thereby supporting subsequent surgical planning. This noninvasive predictive model holds significant potential in clinical application. The CTA-based screening strategy is thus expected to become the preferred evaluation tool for patients with IIH or PT, significantly reducing unnecessary DSA examinations. In the future, we will complete multicenter prospective validation to optimize the generalizability of the model, which can also be deeply integrated with hospital picture archiving and communication systems and artificial intelligence-assisted diagnostic technologies, ultimately enabling the automatic extraction of image features and structured reports.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-157/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-157/dss

Funding: This work was supported by the National Natural Science Foundation of China (grant Nos. 61931013, 82202098, and 82171886), the Natural Science Foundation of Beijing Municipality (grant No. 7222301), the Beijing Key Clinical Discipline Funding (grant No. 2021-135), the Beijing Tongzhou District Science and Technology Plan (grant No. KJ2024CX057), and Beijing Scholar 2015 {grant No. [2015]160}.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-157/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Beijing Friendship Hospital, Capital Medical University (Nos. 2020-P2-202-02 and 2023-P2-095-01), and individual consent for this analysis was waived due to the retrospective nature.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Bi J, Li Z, Zhang X, Bai X, Zhao X, Qu H, Kong Q, An J, Mo D, Sui B. Differentiation Between the Low and High Trans-Stenotic Pressure Gradient in Patients With Idiopathic Intracranial Hypertension Using 4D Flow MRI-Derived Hemodynamic Parameters. J Magn Reson Imaging 2024;59:1569-79. [Crossref] [PubMed]
2. Fargen KM. Idiopathic intracranial hypertension is not idiopathic: proposal for a new nomenclature and patient classification. J Neurointerv Surg 2020;12:110-4.
3. Hsieh YL, Zuo B, Shi Y, Wang S, Wang W. Dynamics of cerebrospinal fluid pressure alterations and bilateral transverse-sigmoid sinus morphologies in Asian patients with venous pulsatile tinnitus. J Int Med Res 2023;51:3000605231187949. [Crossref] [PubMed]
4. Liu KC, Starke RM, Durst CR, Wang TR, Ding D, Crowley RW, Newman SA. Venous sinus stenting for reduction of intracranial pressure in IIH: a prospective pilot study. J Neurosurg 2017;127:1126-33. [Crossref] [PubMed]
5. Abdalkader M, Nguyen TN, Norbash AM, Raz E, Shapiro M, Lenck S, Brinjikji W, Weber P, Sakai O. State of the Art: Venous Causes of Pulsatile Tinnitus and Diagnostic Considerations Guiding Endovascular Therapy. Radiology 2021;300:2-16. [Crossref] [PubMed]
6. Fargen KM, Liu K, Garner RM, Greeneway GP, Wolfe SQ, Crowley RW. Recommendations for the selection and treatment of patients with idiopathic intracranial hypertension for venous sinus stenting. J Neurointerv Surg 2018;10:1203-8. [Crossref] [PubMed]
7. Reid K, Winters HS, Ang T, Parker GD, Halmagyi GM. Transverse Sinus Stenting Reverses Medically Refractory Idiopathic Intracranial Hypertension. Front Ophthalmol (Lausanne) 2022;2:885583. [Crossref] [PubMed]
8. Zhao P, Ding H, Lv H, Li X, Qiu X, Zeng R, Wang G, Wei J, Jin L, Yang Z, Gong S, Wang Z. CT venography correlate of transverse sinus stenosis and venous transstenotic pressure gradient in unilateral pulsatile tinnitus patients with sigmoid sinus wall anomalies. Eur Radiol 2021;31:2896-902. [Crossref] [PubMed]
9. Sunderland K, Wang M, Pandey AS, Gemmete J, Huang Q, Goudge A, Jiang J. Quantitative analysis of flow vortices: differentiation of unruptured and ruptured medium-sized middle cerebral artery aneurysms. Acta Neurochir (Wien) 2021;163:2339-49. [Crossref] [PubMed]
10. Cai L, Zhao E, Niu H, Liu Y, Zhang T, Liu D, Zhang Z, Li J, Qiao P, Lv H, Ren P, Zheng W, Wang Z. A machine learning approach to predict cerebral perfusion status based on internal carotid artery blood flow. Comput Biol Med 2023;164:107264. [Crossref] [PubMed]
11. Dinkin MJ, Patsalides A. Venous Sinus Stenting in Idiopathic Intracranial Hypertension: Results of a Prospective Trial. J Neuroophthalmol 2017;37:113-21. [Crossref] [PubMed]
12. Zhao P, Lv H, Dong C, Niu Y, Xian J, Wang Z. CT evaluation of sigmoid plate dehiscence causing pulsatile tinnitus. Eur Radiol 2016;26:9-14. [Crossref] [PubMed]
13. Dai C, Zhao P, Ding H, Lv H, Qiu X, Li X, Xu N, Meng X, Wang G, Xie J, Yang Z, Gong S, Wang Z. CT evaluation of unilateral pulsatile tinnitus with jugular bulb wall dehiscence. Eur Radiol 2023;33:4464-71. [Crossref] [PubMed]
14. Stoltzfus JC. Logistic regression: a brief primer. Acad Emerg Med 2011;18:1099-104. [Crossref] [PubMed]
15. Schober P, Vetter TR. Logistic Regression in Medical Research. Anesth Analg 2021;132:365-6. [Crossref] [PubMed]
16. Elkahwagy DMAS, Kiriacos CJ, Mansour M. Logistic regression and other statistical tools in diagnostic biomarker studies. Clin Transl Oncol 2024;26:2172-80. [Crossref] [PubMed]
17. Ingwersen EW, Stam WT, Meijs BJV, Roor J, Besselink MG, Groot Koerkamp B, de Hingh IHJT, van Santvoort HC, Stommel MWJ, Daams FDutch Pancreatic Cancer Group. Machine learning versus logistic regression for the prediction of complications after pancreatoduodenectomy. Surgery 2023;174:435-40. [Crossref] [PubMed]
18. Wu T, Wei Y, Wu J, Yi B, Li H. Logistic regression technique is comparable to complex machine learning algorithms in predicting cognitive impairment related to post intensive care syndrome. Sci Rep 2023;13:2485. [Crossref] [PubMed]
19. Lenck S, Vallée F, Labeyrie MA, Touitou V, Saint-Maurice JP, Guillonnet A, Tantot A, Crassard I, Bernat AL, Houdart E. Stenting of the Lateral Sinus in Idiopathic Intracranial Hypertension According to the Type of Stenosis. Neurosurgery 2017;80:393-400. [Crossref] [PubMed]
20. Kumpe DA, Seinfeld J, Huang X, Mei Q, Case DE, Roark CD, Subramanian PS, Lind KE, Pelak VS, Bennett JL. Dural sinus stenting for idiopathic intracranial hypertension: factors associated with hemodynamic failure and management with extended stenting. J Neurointerv Surg 2017;9:867-74. [Crossref] [PubMed]
21. Shazly TA, Jadhav AP, Aghaebrahim A, Ducruet AF, Jankowitz BT, Jovin TG, Bonhomme GR. Venous sinus stenting shortens the duration of medical therapy for increased intracranial pressure secondary to venous sinus stenosis. J Neurointerv Surg 2018;10:310-4. [Crossref] [PubMed]
22. Zhang Y, Ma C, Li C, Li X, Liu R, Liu M, Zhu H, Liang F, Wang Y, Dong K, Jiang C, Miao Z, Mo D. Prediction of the trans-stenotic pressure gradient with arteriography-derived hemodynamic features in patients with idiopathic intracranial hypertension. J Cereb Blood Flow Metab 2022;42:1524-33. [Crossref] [PubMed]
23. Ma C, Zhu H, Liang S, Chang Y, Mo D, Jiang C, Zhang Y. Prediction of Venous Trans-Stenotic Pressure Gradient Using Shape Features Derived From Magnetic Resonance Venography in Idiopathic Intracranial Hypertension Patients. Korean J Radiol 2024;25:74-85. [Crossref] [PubMed]
24. West JL, Greeneway GP, Garner RM, Aschenbrenner CA, Singh J, Wolfe SQ, Fargen KM. Correlation between angiographic stenosis and physiologic venous sinus outflow obstruction in idiopathic intracranial hypertension. J Neurointerv Surg 2019;11:90-4. [Crossref] [PubMed]
25. Li X, Qiu X, Ding H, Lv H, Zhao P, Yang Z, Gong S, Wang Z. Effects of different morphologic abnormalities on hemodynamics in patients with venous pulsatile tinnitus: A four-dimensional flow magnetic resonance imaging study. J Magn Reson Imaging 2021;53:1744-51. [Crossref] [PubMed]
26. Yang L, Zhang D. Case report: Coexistence of Labbe vein thrombosis and autoimmune encephalitis with two different antibodies. Front Neurol 2023;14:1170169. [Crossref] [PubMed]