Prediction of brain cancer-associated thrombosis by imaging-based machine learning: Results from the Vienna Cancer and Thrombosis Study

Introduction Patients with primary brain cancers are at high risk of venous thromboembolism (VTE), including pulmonary embolism and deep vein thrombosis. Risk prediction of VTE might be improved by index cranial MRI features. Method We utilized data from patients with newly diagnosed or recurrent primary brain tumors of WHO grades 3 and 4 included into the Vienna Cancer and Thrombosis Study (Vienna-CATS), a prospective cohort study conducted between 2003 and 2019. The primary outcome was VTE during a follow-up period of 2 years. Cranial MRI scans performed within 6 weeks before (preoperative MRI) and up to 3 days after (postoperative MRI) the diagnostic tumor resection or biopsy were analyzed. The following MRI imaging features were reviewed and annotated by 2 radiologists on all available images: amount of tumor lesions, lesion size, minimal and mean Apparent Diffusion Coefficient (ADC), intratumoral hemorrhage, contrast enhancement, necrosis, perifocal edema, chronic small vessel ischemia/white matter lesions, and extent of resection (only on postoperative MRI). The annotated features were combined with the parameters age, sex, and WHO grading and entered into a dataset, which was used to train and validate 6 classification models (Decision Tree [DT], Logistic Regression [LGR], K-Nearest Neighbor [KNN], Extreme Gradient Boosting [XGB], Random Forest [RF], and Explainable Boosting Machine [EBM]) for VTE prediction at 6 months and 2 years. Predictive performance was evaluated using Monte Carlo cross-validation. Preprocessing algorithms were fitted on the training data and applied to the test data to avoid data leakage. Discrimination and input parameter influence were assessed. Results In total, 230 patients were included. Of those, 83 (36.1%) were female and the median age (interquartile range) was 58 (45-67) years ([ Fig. 2 ]). At 6 months and 2 years, 20 and 28 VTE occurred, respectively. Based on the number of available cranial MRI scans (223 preoperative and 79 postoperative), only the preoperative scans were used for the machine learning analysis. Furthermore, as not all required MRI sequences were available consistently across the study duration, only features listed in Fig. 2 entered the analysis. The models’ area under the receiver operating characteristics (AUC, 95% confidence interval) ranged from 0.53 (0.52-0.54) for K-Nearest Neighbor at 2 years to 0.65 (0.64-0.67) for Explainable Boosting Machine at 6 months, indicating at best moderate performance. The input feature relevance was evaluated individually, revealing lesion size and presence of contrast enhancement as the most important imaging parameters ([ Fig. 1 ]). However, compared to the non-imaging parameters age, sex, and WHO grading, the imaging features added only modest predictive value. Fig. 2 Lesion size is the product of the two largest lesion diameters in the plain with the largest visible tumor mass. The sum of all lesions was calculated if more than one lesion was detected. BMI, body mass index; MRI, magnetic resonance imaging; VTE, venous thromboembolism; WHO, world health organization. Conclusion Imaging features extracted from cranial MRI obtained during the diagnostic work-up of brain cancer contributed to VTE prediction, although their overall contribution was limited. Fig. 1 Permutation importance of the 6-months explainable boosting machine (EBM) model (upper panel) and the 2-year logistic regression (LGR) model (lower panel). These models were selected because they showed the best performance at the respective time point. Wider bars represent higher importance. Black bars represent standard deviation. WHO, World Health Organization. Publication History Article published online: 17 February 2026 © 2026. Thieme. All rights reserved. Georg Thieme Verlag KG Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany