Predictive models for trans-stenotic pressure gradient in the transverse sinus based on multimodal correlates Authors: Yan Huang, Linkun Cai, Xiaoyu Qiu, Chihang Dai, Xu Han, Guo-Peng Wang, Shusheng Gong, Zhenghan Yang, Heyu Ding, Zhenchang Wang, Pengfei Zhao Published: 2026 ABSTRACT BackgroundThe invasive nature of the methods currently available for screening based on the trans-stenotic pressure gradient (TPG) limits their widespread application. We aimed to develop machine learning models for predicting TPG in the transverse sinus.MethodsThe study was approved by the Ethics Committee of Beijing Friendship Hospital Affiliated to Capital Medical University. We retrospectively analyzed the data of 83 patients confirmed to have pressure-measured transverse sinus stenosis from January 2017 to June 2022 at Beijing Friendship Hospital. The dataset included clinical, laboratory, computed tomography angiography, and four-dimensional flow magnetic resonance data with 29 feature indicators. Six machine learning methods—support vector machine, k-nearest neighbors, decision tree, random forest (RF), gradient boosting decision tree, and extreme gradient boosting—were employed to construct classification models for 4- and 8-mmHg TPG thresholds via fivefold cross-validation. Models were evaluated according to accuracy, sensitivity, specificity, F1 score, the Matthews correlation coefficient (MCC), and area under the curve (AUC).ResultsFor the 4-mmHg threshold, RF showed the highest performance, with an accuracy of 0.83, a specificity of 0.62, a sensitivity of 0.95, an F1 score of 0.92, an MCC of 0.60, and an AUC of 0.92. RF also excelled for the 8-mmHg threshold, with accuracy, specificity, sensitivity, F1 score, MCC, and AUC values of 0.84, 0.86, 0.81, 0.78, 0.70, and 0.91, respectively. SHapley Additive exPlanations analysis indicated that peak velocity and average velocity differences positively correlated with TPG, whereas the residual area ratio was negatively correlated.ConclusionsThe models demonstrated strong discriminative ability and are useful for noninvasive TPG prediction. Introduction The trans-stenotic pressure gradient (TPG) indicates the severity of blood flow obstruction caused by vessel stenosis. The TPG of the transverse sinus is a critical factor in conditions such as idiopathic intracranial hypertension (IIH) and pulsatile tinnitus (PT) (1,2). Venous sinus stenting (VSS) has been proven to be an effective treatment for patients with IIH or PT. A TPG exceeding 8 mmHg is commonly used as an indication for stenting (3-5), as the pressure gradient from the superior sagittal sinus to the superior vena cava usually does not exceed 4 mmHg in normal conditions (6). Digital subtraction angiography (DSA) combined with microcatheter-based direct pressure measurement is currently the gold standard for assessing the TPG of the transverse sinus. However, the invasive nature of this and similar methods limits their widespread application in TPG-based screening (7). The noninvasive evaluation of the transverse sinus TPG could provide more precise guidance for selecting appropriate candidates for DSA pressure measurements. Several predictive models have been proposed, with the transverse sinus being a prominent factor. West et al. (8) found that for every 10% increase in transverse sinus stenosis, the intrasinus pressure increases by approximately 3.4 mmHg. Zhao et al. (9) demonstrated a linear positive correlation between the length and severity of stenosis in the affected transverse sinus and the TPG through computed tomography angiography (CTA). Furthermore, studies have shown that flow velocity derived from four-dimensional (4D) flow magnetic resonance (MR) and jet-like blood flow patterns in the transverse sinus are indicative of symptomatic stenosis of the transverse sinus (10), laying the groundwork for the noninvasive assessment of transverse sinus TPG. Machine learning has been widely used in the evaluation of various types of diseases, with information from different modalities being integrated to construct predictive models. For instance, Guédon et al. (11) integrated classification and regression trees algorithms with genetic algorithms to develop an effective prediction model for outcomes of flow-diverting stent treatment of intracranial aneurysm, achieving outstanding results. Berhane et al. (12) successfully combined 4D flow MR data with machine learning techniques to construct a predictive model for intracranial aneurysm hemodynamics. Xia et al. (13)employed multivariate logistic regression analysis and random forest (RF) algorithms to develop a risk prediction model for intracranial aneurysm rupture, accurately predicting patient outcomes. These studies collectively highlight the efficacy of integrating machine learning algorithms with imaging data to characterize the complex pathogenesis of cardiovascular diseases and construct predictive models. Therefore, combining machine learning algorithms with imaging and clinical data represents a promising approach for developing a comprehensive TPG evaluation model. In this study, we analyzed patients who underwent transverse sinus TPG-based screening to identify potential clinical, laboratory, and imaging indicators. Furthermore, we used machine learning methods in conjunction with multiparameter data to construct noninvasive models for predicting the transverse sinus TPG. This study aimed to enhance our understanding of transverse sinus TPG and provide objective guidance for selecting candidates suitable for venous manometry. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0646/rc). Methods Participants This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and approved by the Ethics Committee of Beijing Friendship Hospital, Capital Medical University (No. 2023-P2-095-01). Due to the retrospective design, the requirement for informed consent was waived. We retrospectively collected data on patients who underwent DSA imaging with venous manometry at our center from January 2017 to June 2022 for various medical conditions. The inclusion criteria were (I) transverse sinus stenosis and (II) completion of CTA and 4D flow MR examinations. Meanwhile, the exclusion criteria were as follows: (I) imaging findings showing aneurysms, arteriovenous malformations, dural arteriovenous fistulas, or other abnormal vascular communication; (II) ear and intracranial inflammatory or neoplastic lesions; (III) other conditions affecting intracranial hemodynamics; and (IV) incomplete laboratory examinations. Four-dimensional flow MR image acquisition and data measurement Both 2D phase-contrast (PC) and 4D flow MR imaging sequences were acquired with retrospective electrocardiogram gating. The velocity encoding for 2D PC MR was initially set to 80 cm/s. If velocity aliasing was observed, the encoding was incrementally increased by 20 cm/s (to 100, 120, and 140 cm/s) until aliasing was no longer present. The maximum nonaliased velocity encoding value (VENC) determined on 2D PC MRI was multiplied by 1.2 to set the VENC for 4D flow MR. For example, if 100 cm/s showed no aliasing on 2D PC, the 4D flow VENC was set to 120 cm/s. The actual VENC ranges used for all patients were as follows: 2D PC, 80–140 cm/s; and 4D flow, 96–168 cm/s. The detailed acquisition parameters are provided in Table S1. GTFlow 2.2.15 software (GyroTools, Zurich, Switzerland) was used for preprocessing, visualization, quantification, evaluation, and interpretation of the 4D flow data. Preprocessing steps included region masking, eddy current correction, velocity aliasing correction, and vessel segmentation. The measurement planes for upstream and downstream velocities were positioned at the nearest normal-caliber segments immediately proximal and distal to the stenosis, respectively (within 5–10 mm from the stenotic lesion), with each plane placed perpendicular to the vessel centerline. Vessel contours were manually delineated on the cross-sectional images to define the region of interest (ROI). The software automatically calculated the mean velocity as the average over all pixels within the ROI and the peak velocity as the maximum pixel value within the ROI. Two independent neuroradiologists, each with 5 years of experience, performed all measurements; both were completely blinded to the invasive TPG results. The average and peak velocities downstream and upstream of the transverse sinus stenosis were measured, and their differences were calculated. The measurement protocols followed those used in our previous studies (14-16). CTA image acquisition and measurement For CTA, we used either a Brilliance 64-slice spiral CT scanner (Philips, Amsterdam, the Netherlands) or a Revolution 256-slice spiral CT scanner (Siemens Healthineers, Erlangen, Germany). Of the 83 patients, 37 were scanned on the Brilliance 64-slice CT scanner and 46 on the Revolution 256-slice CT scanner. Both scanners included a reconstruction slice thickness of 0.625 mm with a reconstruction interval of 0.3 mm. The scanning range extended from the sixth cervical vertebra to the cranial vertex. The scan parameters were as follows: tube voltage, 100 kV; tube current, 300 mA; rotation time, 0.75 s per rotation; matrix size, 512×512; and detector arrangement, 64×0.625 mm. Iopamidol (370 mg/mL) was intravenously administered as a contrast agent through the antecubital vein at a dose of 1.5 mL/kg, with an injection rate of 5 mL/s. Following the contrast injection, 20 mL of normal saline was injected at the same rate. Bolus-tracking technology was employed for automated triggering, with the trigger point set at the ascending aorta, a trigger threshold of 150 HU, and an ROI of 200 mm2. The scan was performed from the cranial vertex to the cranial base with a delay of 35–40 s. CTA images in Digital Imaging and Communications on Medicine Format were uploaded to the Extended Brilliance Workspace postprocessing workstation (Philips). The window width and level were set to 700 and 200 HU, respectively. Measurements included the residual area ratio (the area at the site of stenosis divided by the area downstream from the stenosis) and the length of the stenosis. The assessment also included the type of stenosis, location of the Labbé vein (categorized as upper, middle, or lower), and the type of drainage dominance (calculated as the cross-sectional area of the symptomatic side divided by the sum of the symptomatic and contralateral side areas). The degree of contralateral stenosis was also assessed. The criteria for assessing each indicator were obtained from previous studies by Zhao et al. (9,17). DSA and venous manometry DSA examinations were performed within 7 days of image acquisition. After local anesthesia was administered, the Seldinger technique was employed for intravascular DSA pressure measurements with an angiography machine (Innova 4100-IQ, GE HealthCare, Chicago, IL, USA). The patient was positioned supine. First, a percutaneous puncture was made in the right femoral artery. A 5F catheter was then carefully guided into the left and right internal carotid arteries and the left vertebral artery for angiography, allowing observation of the distribution, course, and morphology of the intracranial arteries and veins. Subsequently, a vascular sheath was inserted through the percutaneous puncture in the right femoral vein. A 5F catheter was advanced into the internal jugular vein of the affected side, followed by superselection via a 2.7F microcatheter inserted into the intracranial venous sinus. The pressure sensor was connected to the microcatheter, which was positioned at the distal end of the transverse sinus stenosis on the affected side and at the proximal end of the sigmoid sinus. All pressure measurements were performed with the patient breathing freely under local anesthesia. The transducer was zeroed at the level of the right atrium (mid-axillary line). At each measurement site, the pressure recording was maintained for at least 30 seconds, and the mean pressure over three consecutive cardiac cycles was recorded. Each site was measured twice, and the final TPG was calculated as the average of the two measurements. All measurements were performed by an interventional radiologist with 10 years of experience. The TPG of the transverse sinus was then measured. A schematic diagram of the pressure measurement sites is provided in Figure S1. Collection of clinical data and laboratory indicators Baseline patient information, including sex, age, and body mass index, was collected. Additionally, biochemical indicators, such as blood glucose, total cholesterol, triglycerides, low-density lipoprotein cholesterol, and high-density lipoprotein cholesterol, were recorded. Routine blood parameters (red blood cells, hemoglobin, and platelets), thyroid series (thyroid-stimulating hormone, thyroid uptake rate, thyroid hormone, triiodothyronine, and thyroxine), disseminated intravascular coagulation screening (fibrinogen and D-dimer), and endocrine parameters (estradiol, luteinizing hormone, follicle-stimulating hormone, testosterone, prolactin, and progesterone) were also assessed. Machine learning model construction The TPG of the transverse sinus measured by DSA was used to define two binary classification tasks based on thresholds of 4 and 8 mmHg, respectively. Candidate predictors included imaging-derived and laboratory variables, as detailed in Table S2. Separate prediction models were developed for each classification task. To prevent data leakage, all data-dependent preprocessing and feature-selection procedures were incorporated into a nested stratified fivefold cross-validation framework and performed independently for each classification task. This nested cross-validation framework was used to evaluate model generalization and stability, ensuring that the models could be effectively extended to new data (18). The complete dataset was divided into five stratified outer folds. In each outer iteration, four folds were used as the outer training set, while the remaining fold was retained as an untouched outer evaluation fold. Within each outer training set, feature preprocessing and least absolute shrinkage and selection operator (LASSO)-based feature selection were conducted with a fivefold inner cross-validation procedure. The LASSO regularization parameter was optimized exclusively within the inner cross-validation loop. The normalization parameters and selected feature subset derived from the outer training set were subsequently applied without modification to the corresponding held-out outer evaluation fold. Except for the LASSO regularization parameter, all classifier hyperparameters were fixed a priori according to the values listed in Table 1. Six machine learning classifiers were evaluated: support vector machine (SVM), k-nearest neighbors (KNN), decision tree (DT), RF, gradient boosting decision tree (GBDT), and extreme gradient boosting (XGBoost). All machine learning models were implemented using Python 3.9.13 (Python Software Foundation, Wilmington, DE, USA), and all experiments were run on a single GeForce RTX 3090 GPU (Nvidia, Santa Clara, CA, USA) with 24 GB of memory. The complete machine learning workflow is illustrated in Figure 1. Table 1Hyperparameter values of the machine learning modelsModelHyperparameter valueSVMC = 1.0, Kernel = rbf, gamma = scaleKNNN_neighbor =5, leaf_size =30DTCriterion = gini, min_samples_split =2, min_samples_leaf =1RFCriterion = gini, N_estimators =100, min_samples_split =2, min_samples_leaf =1GBDTN_estimators =100, learning_rate =0.1, max_depth =3, min_samples_split =2, min_samples_leaf =1XGboostN_estimators =100, learning_rate =0.1, max_depth =3, min_child_weight =1DT, decision tree; GBDT, gradient boosting decision tree; KNN, k-nearest neighbors; RF, random forest; SVM, support vector machine; XGBoost, extreme gradient boosting. Hyperparameter values of the machine learning models DT, decision tree; GBDT, gradient boosting decision tree; KNN, k-nearest neighbors; RF, random forest; SVM, support vector machine; XGBoost, extreme gradient boosting. Process of machine learning model construction. DT, decision tree; GBDT, gradient boosting decision tree; KNN, k-nearest neighbors; RF, random forest; SVM, support vector machine; TPG, trans-stenotic pressure gradient; XGBoost, extreme gradient boosting. Statistical analysis In the feature-screening process, features were removed based on a correlation analysis, with a P value greater than 0.05 and features with an eigenvalue weight of 0 included for LASSO regression. To evaluate the performance of the machine learning models, we employed six metrics: accuracy, sensitivity, specificity, F1 score, and Matthews correlation coefficient (MCC). Accuracy represents the ratio of correctly predicted samples to the total number of samples. Sensitivity measures the model's ability to correctly identify positive class samples and is crucial for detecting true positives. Specificity quantifies the model's ability to correctly identify negative class samples and is critical to minimizing false positives. F1 score combines precision and recall (sensitivity) to balance the tradeoff between them. The MCC reflects the relationship between true positives, true negatives, false positives, and false negatives, providing a comprehensive measure of model performance. Receiver operating characteristic (ROC) curves were generated via the continuous prediction scores obtained from each classifier. Predictions from the held-out samples across the fivefold cross-validation procedure were aggregated to generate pooled out-of-fold predictions for area under the curve (AUC) analysis. Finally, we employed game-theory-based SHapley Additive exPlanations (SHAP) analysis to evaluate the impact of each input variable on the prediction results from both macro- (global) and micro- (local) perspectives. Results Participants’ characteristics This study included data from 83 patients (12 males and 71 females), aged 22 to 74 years (mean 38.9±11.8 years). The transverse sinus TPG ranged from 1.03 to 16.99 mmHg (mean 7.84±5.15 mmHg). Specifically, 21 patients had a TPG less than 4 mmHg, whereas 62 patients had a TPG greater than 4 mmHg. Additionally, 50 patients had a TPG less than 8 mmHg, and 33 patients had a TPG greater than 8 mmHg. The basic information and examination results of all patients are summarized in Table 2. Table 2Demographic characteristics of patientsCharacteristicsRange/n (%)Mean ± SDAge (years)22–7438.9±11.8SexNA Males12 (14.46)NA Females71 (85.54)NATPG1.03–16.997.84±5.15 ≥4 mmHg62 (74.70)NA <4 mmHg21 (25.30)NA ≥8 mmHg33 (39.76)NA <8 mmHg50 (60.24)NANA, not applicable; SD, standard deviation; TPG, trans-stenotic pressure gradient. Demographic characteristics of patients NA, not applicable; SD, standard deviation; TPG, trans-stenotic pressure gradient. Feature screening In the LASSO regression analysis, three characteristics had a weight value of 0: free triiodothyronine, free thyroxine, and fibrinogen levels. In the correlation analysis, nine features were significantly correlated with TPG (P<0.05), including average velocity difference, peak velocity difference, residual area ratio, stenosis length, location of the Labbé vein, age, thyroxine, luteinizing hormone, and follicle-stimulating hormone levels. Finally, nine features were selected for model construction based on the intersection of the two analyses, as shown in Table 3. Table 3The features included in model constructionFeatureTypeAverage velocity differenceMRPeak velocity differenceResidual area ratioCTStenosis lengthLocation of the Labbé veinAgeNonradiographic featuresThyroxineLuteinizing hormoneFollicle-stimulating hormoneCT, computed tomography; MR, magnetic resonance. The features included in model construction CT, computed tomography; MR, magnetic resonance. Performance of machine learning models To determine the optimal model for predicting the transverse sinus TPG, we conducted extensive experiments to validate the performance of each machine learning model. As shown in Table 4, in the binary classification model based on a 4-mmHg threshold, the DT model achieved the highest accuracy score (0.86), followed by KNN (0.84), RF (0.83), XGBoost (0.83), GBDT (0.82), and SVM (0.77). This indicates that the machine learning models can effectively distinguish transverse sinus TPG. The DT model performed well, with an accuracy of 0.86, a specificity of 0.96, and an MCC of 0.67. However, the RF model outperformed the DT model in terms of the F1 score. The summarized experimental results are highlighted in bold, with the best outcome for each evaluation metric. The ROC curves, shown in Figure 2, reflect the prediction results of each classifier, with the average sensitivity and specificity being combined. The AUC represents classifier performance, with a larger AUC indicating better performance. The values shown in brackets represent the minimum and maximum AUC values among the five validation folds. The AUC values for SVM, KNN, DT, GBDT, and XGBoost were 0.87, 0.80, 0.77, 0.88, and 0.84, respectively. We selected RF as the optimal model as it had the highest AUC (0.92) although DT yielded a higher accuracy and MCC. Table 4Performance of the machine learning models for binary classification based on a 4-mmHg thresholdModelAccuracySpecificitySensitivityF1 scoreMCCSVM0.77±0.120.58±0.210.84±0.160.84±0.100.44±0.25KNN0.84 ±0.030.43±0.100.98±0.030.90±0.020.56±0.09DT0.86±0.030.77±0.200.87±0.040.89±0.010.67±0.07RF0.83±0.070.62±0.250.95±0.090.92±0.030.60±0.10GBDT0.82±0.060.72±0.170.89±0.110.89±0.040.59±0.09XGBoost0.83±0.070.62±0.250.91±0.120.89±0.050.59±0.15Data are presented as mean ± SD. DT, decision tree; GBDT, gradient boosting decision tree; KNN, k-nearest neighbors; MCC, Matthews correlation coefficient; RF, random forest; SD, standard deviation; SVM, support vector machine; XGBoost, extreme gradient boosting. Performance of the machine learning models for binary classification based on a 4-mmHg threshold Data are presented as mean ± SD. DT, decision tree; GBDT, gradient boosting decision tree; KNN, k-nearest neighbors; MCC, Matthews correlation coefficient; RF, random forest; SD, standard deviation; SVM, support vector machine; XGBoost, extreme gradient boosting. Receiver operating characteristic curve of each machine learning method for binary classification based on a 4-mmHg threshold. Data in brackets represent the range (minimum–maximum) of AUC values across the fivefold cross-validation. AUC, area under the curve; CV, cross validation; DT, decision tree; GBDT, gradient boosting decision tree; KNN, k-nearest neighbors; RF, random forest; SVM, support vector machine; XGB (XGBoost), extreme gradient boosting. In the binary classification model based on an 8-mmHg threshold, the RF model achieved the highest accuracy score (0.84), followed by the GBDT (0.82), XGBoost (0.77), SVM (0.77), DT (0.75), and KNN (0.74) models. This indicates that the machine learning model also performed well in distinguishing an 8-mmHg TPG. The RF model demonstrated strong performance with an accuracy of 0.84, a specificity of 0.86, and an MCC of 0.70. Notably, the KNN model achieved the highest sensitivity (0.90) among all models. The detailed results are presented in Table 5, with the summarized experimental results highlighted in bold, showcasing the best outcomes for each evaluation metric. To compare the diagnostic efficacy of different models for an 8-mmHg TPG, we analyzed the ROC curves. The AUC values of SVM, KNN, DT, RF, GBDT, and XGBoost were 0.87, 0.86, 0.71, 0.91, 0.89, and 0.87, respectively, with RF having the largest AUC and the strongest diagnostic efficacy (Figure 3). Table 5Performance of machine learning models for binary classification based on a 8-mmHg thresholdModelAccuracySpecificitySensitivityF1 scoreMCCSVM0.77±0.060.76±0.150.79±0.120.74±0.050.56±0.11KNN0.74±0.070.62±0.150.90±0.200.72±0.100.54±0.14DT0.75±0.070.76±0.190.66±0.180.66±0.110.51±0.12RF0.84±0.080.86±0.190.81±0.120.78±0.070.70±0.12GBDT0.82±0.040.82±0.120.81±0.120.78±0.040.64±0.05XGBOOST0.77±0.040.76±0.150.79±0.150.73±0.030.56±0.04Data are presented as mean ± SD. DT, decision tree; GBDT, gradient boosting decision tree; KNN, k-nearest neighbors; MCC, Matthews correlation coefficient; RF, random forest; SD, standard deviation; SVM, support vector machine; XGBoost, extreme gradient boosting. Performance of machine learning models for binary classification based on a 8-mmHg threshold Data are presented as mean ± SD. DT, decision tree; GBDT, gradient boosting decision tree; KNN, k-nearest neighbors; MCC, Matthews correlation coefficient; RF, random forest; SD, standard deviation; SVM, support vector machine; XGBoost, extreme gradient boosting. Receiver operating characteristic curve of each machine learning method for binary classification based on a 8-mmHg threshold. Data in brackets represent the range (minimum–maximum) of AUC values across the fivefold cross-validation. AUC, area under the curve; CV, cross validation; DT, decision tree; GBDT, gradient boosting decision tree; KNN, k-nearest neighbors; RF, random forest; SVM, support vector machine; XGB (XGBoost), extreme gradient boosting. Interpretability analysis In the binary classification model based on a 4-mmHg threshold, the RF model showed the highest AUC. Consequently, we analyzed the interpretability of the RF model. As shown in Figure 4A, SHAP analysis identified the most relevant features for the RF model: peak velocity difference, residual area ratio, and average velocity difference. Subsequently, we analyzed the positive and negative relationships between these features and TPG. Among these features, the peak velocity difference and average velocity difference exhibited a positive correlation with TPG, whereas the residual area ratio showed a negative correlation (Figure 4B). SHAP analysis of a binary classification model based on a 4-mmHg threshold. (A) Bar chart showing the feature importance ranking based on mean absolute SHAP values. (B) SHAP summary plot generated from the random forest model. The horizontal axis represents SHAP values, and each point represents an individual sample. The color indicates the feature value, with red representing higher values and blue representing lower values. Positive SHAP values indicate that the corresponding feature contributes to increasing the model output toward the positive class (higher TPG category), whereas negative SHAP values indicate the opposite effect. Features are ordered according to their model-specific mean absolute SHAP values, with higher-ranked features indicating greater overall contribution to model predictions. SHAP, SHapley Additive exPlanations. In the binary model based on an 8-mmHg threshold, the RF model exhibited the highest accuracy, specificity, MCC, and AUC. Therefore, we considered the RF model to be the best-performing model. Similarly, we identified the features associated with the RF model for classification using the 8-mmHg threshold. Among these features, the three most relevant ones were peak velocity difference, residual area ratio, and average velocity difference, as shown in Figure 5A. Figure 5B illustrates that the peak velocity difference and average velocity difference exhibited a positive correlation with TPG, whereas the residual area ratio demonstrated a negative correlation. SHAP analysis of a binary classification model based on an 8-mmHg threshold. (A) Bar chart showing the feature importance ranking based on mean absolute SHAP values. (B) SHAP summary plot generated from the random forest model. The horizontal axis represents SHAP values, and each point represents an individual sample. The color indicates the feature value, with red representing higher values and blue representing lower values. Positive SHAP values indicate that the corresponding feature contributes to increasing the model output toward the positive class (higher TPG category), whereas negative SHAP values indicate the opposite effect. Features are ordered according to their model-specific mean absolute SHAP values, with higher-ranked features indicating greater overall contribution to model predictions. SHAP, SHapley Additive exPlanations. Discussion This study used data from CTA, 4D flow MR, and laboratory examinations. A total of 29 feature indicators were collected, with elevated TPG due to transverse sinus stenosis serving as the dependent variable. We constructed classification models using six different machine learning methods for thresholds of 4 and 8 mmHg. The binary classification models demonstrated strong performance, with the RF model performing optimally. Interpretability analysis revealed that the peak velocity difference, residual area ratio, and average velocity difference were the most important features in the model. Additionally, both the peak velocity difference and average velocity difference were positively correlated with TPG. Combining machine learning algorithms with imaging data is an effective approach for elucidating the complex pathogenesis of cardiovascular diseases and constructing predictive models. This approach is particularly advantageous for diseases characterized by multiple intrinsic factors and interrelated features (19,20). In our study, we employed commonly used machine learning models, including SVM, KNN, and DT, as comparative models and RF, GBDT, and XGBoost as ensemble models. Notably, regardless of whether the binary classification models were based on a threshold of 4 or 8 mmHg, the ensemble models consistently outperformed the individual classifiers. This superiority can be attributed to the ensemble learning algorithms’ ability to reduce the variance caused by considering a single decision tree for the same dataset (21). VSS has shown favorable outcomes for patients with IIH, including a headache relief rate of 78–83%, an optic disc edema resolution rate of 87–97%, a visual symptom improvement rate of 74–85%, and a tinnitus relief rate of 95% (22-24). Currently, no universally established optimal threshold of venous sinus TPG during VSS in patients with IIH exists. A systematic review suggested that a TPG range of 6–8 mmHg is a reasonable criterion for VSS (3). Zhang et al. (25) developed machine learning models based on TPG values of 6 and 8 mmHg to predict the TPG and achieved promising results. Additionally, some studies have proposed an operative threshold based on a TPG greater than 8 mmHg (26,27). Therefore, if the transverse sinus TPG exceeds 8 mmHg, stent implantation remains a viable option, which is why we adopted 8 mmHg as the binary classification threshold. Furthermore, considering that the maximum pressure gradient between the superior sagittal sinus and superior vena cava typically does not exceed 4 mmHg (6), we deemed transverse sinus narrowing below 4 mmHg as normal and unnecessary for stent placement. Thus, we set 4 mmHg as an additional binary threshold. Our results indicated that both binary classification models performed well and successfully predicted TPG levels. However, our triple-classification models performed poorly, with most achieving an accuracy below 0.6. This limitation may be attributed to our small sample size, resulting in insufficient data volumes for each category and hindering the construction of accurate machine learning models. The interpretability analysis in this study revealed that the transverse sinus TPG was influenced by a complexity of interactions among several features. Notably, the peak and average velocity differences were positively correlated with a TPG value associated with transverse sinus narrowing. These crucial hemodynamic indicators are derived from 4D flow MR measurements, which offer temporal resolution and enable the visualization of blood flow variations throughout the cardiac cycle. Additionally, 4D flow MR quantifies various hemodynamic parameters, including blood flow velocity, flow rate, and intracavitary pressure (28). Ding et al. (29) found a significant positive correlation between TPG and the difference in blood flow velocity measured upstream and downstream of the stenosis using 4D flow MR in patients with transverse sinus narrowing, consistent with our findings. Other studies have reported a high-velocity jet flow occurring after blood passes through a narrowed region (30,31). Guédon et al. used the velocity ratio (highest velocity measured at the sinus stenosis compared to the highest velocity measured at the transverse sinus origin) as an indicator for VSS. In a cohort of 41 patients selected based on blood flow velocity criteria, approximately 97% experienced symptom relief after surgery, with nearly all relevant clinical signs of IIH resolving (32). Consequently, hemodynamic indices at the site of venous sinus stenosis are highly valuable for estimating the TPG and identifying patients who may benefit from further invasive venous manometry, with 4D flow MR providing critical biomechanical information through noninvasive assessment. This study involved several limitations that should be addressed. First, to ensure that all patients had complete CTA and 4D flow MR data, we excluded several patients, resulting in a relatively small number of patients with transverse sinus stenosis who underwent DSA, potentially introducing selection bias. The relatively small sample size might have led to instability in the features selected across training folds and increased the variance of model-performance estimates. Second, the data used in this study were solely obtained from a single center and lacked external validation in other centers, which may impact the model's generalizability. Therefore, the findings should be regarded as internally validated and exploratory until they are confirmed in a larger, independent, and preferably multicenter dataset. Third, CTA examinations were acquired with two different CT scanner models. Differences in scanner hardware, acquisition and reconstruction parameters, spatial resolution, image noise, and contrast enhancement might have affected CTA-derived morphologic measurements despite the use of a consistent image-analysis procedure. Future studies should assess the interobserver and intraobserver reproducibility of key imaging-derived predictors (e.g., average/peak velocity difference and residual area ratio) using repeated blinded measurements and intraclass correlation coefficients. Finally, a few of the laboratory indicators included in our study were selected based on our prior experience. In future research, a more thorough screening of relevant indicators will be necessary to refine the model. Conclusions We developed a machine learning binary classification model for predicting TPG based on multimodal imaging and laboratory data. The model exhibited high discriminatory power, and the interpretability analysis highlighted the relevant features. Our machine learning framework facilitates noninvasive assessment of TPG in the transverse sinus, providing a noninvasive screening tool that can aid in selecting appropriate candidates for invasive DSA pressure measurements. Supplementary The article’s supplementary files as