A logistic regression model combining clinical and CT texture features predicted deep vein thrombosis stability with an AUC of 0.87 (95% CI 0.73-0.87) and an accuracy of 0.79.
Observational (n=108)
Can machine learning models based on clinical and CT texture features accurately predict DVT stability?
Machine learning models, particularly logistic regression, combining clinical and CT texture features can effectively predict DVT stability and the associated risk of acute pulmonary embolism.
Effect estimate: AUC 0.87 (95% CI 0.73-0.87)
Background This study aimed to develop multiple machine learning (ML) models to predict DVT stability based on clinical and computed tomography (CT) texture features. Methods A total of 108 patients diagnosed with DVT by clinical examination and ultrasonography in this study. Patients were divided into the DVT with acute pulmonary embolism (APE) (thrombus unstable group) and DVT without APE (thrombus stable group) groups based on whether their computed tomography pulmonary angiography examination was combined with APE. The region of interest was manually delineated on the CT images using the 3D-Slicer software to extract the textural features of the thrombus. The patients were divided into training and validation sets in a ratio of 7:3. The least absolute shrinkage and selection operator and ten-fold cross-validation were applied to obtain texture features with nonzero coefficients in the training set. Clinical data were used as variables to screen for independent risk factors predicting DVT stability using univariate and multivariate logistic regression analyses. Four machine learning algorithms, logistic regression (LR), support vector machine (SVM), K-nearest neighbor (KNN), and extreme gradient boosting (XGBooST), were used to develop a DVT stability prediction model based on a combination of nonzero feature parameters and clinical features. The performance of the models was assessed and compared using the accuracy, precision, recall, F1 score, specificity, positive prediction rate, negative prediction rate, and area under the curve (AUC), calibration curves, and decision curves. Results The combined AUC, calibration curve, decision curve, and other evaluation metrics showed that the LR model outperformed other ML models AUC: 0.87 (0.73∼0.87), Accuracy: 0.79, Precision: 0.75, F1 Score: 0.77, Recall: 0.80, Specificity: 0.87, Probability of Positive Prediction: 0.82, Probability of Negative Prediction: 0.75, with the best prediction performance. Conclusions ML models based on clinical and CT texture features can be used to predict DVT stability.
Yu et al. (2026) conducted an observational in Deep vein thrombosis (n=108). Machine learning models (Logistic Regression, SVM, KNN, XGBooST) vs. Comparison among models was evaluated on DVT stability prediction performance (AUC) (AUC 0.87, 95% CI 0.73-0.87). A logistic regression model combining clinical and CT texture features predicted deep vein thrombosis stability with an AUC of 0.87 (95% CI 0.73-0.87) and an accuracy of 0.79.