AI-Based Prediction Model for Iliofemoral DVT Thrombolysis

This prospective single-arm cohort study aims to develop an AI-powered prediction model for treatment outcomes in patients with acute extensive iliofemoral deep vein thrombosis (IF-DVT) undergoing stent-free pharmacomechanical thrombolysis. The study addresses the current lack of validated tools for patient selection and outcome prediction in catheter-directed interventions for proximal DVT. Thirty consecutive adult patients with MRV-confirmed acute IF-DVT will undergo pharmacomechanical thrombolysis using the AngioJet ZelanteDVT system with adjunctive rtPA administration. The primary objective is to develop a convolutional neural network (CNN) trained on serial MRV imaging data to predict three-month venous recanalization success. MRV acquisitions occur at baseline, predischarge, and three-month follow-up. Ground truth segmentation will be performed by an experienced radiologist using 3D Slicer, with semi-automated propagation across the dataset. Feature extraction will include geometric metrics, radiomic texture analysis, and morphological characteristics of both thrombus and vessel architecture. Secondary endpoints include acute kidney injury incidence (a significant concern with rheolytic thrombectomy due to hemolysis-induced nephrotoxicity), post-thrombotic syndrome development assessed via Villalta scoring, and various safety outcomes including major bleeding per ISTH criteria. The study protocol incorporates rigorous monitoring for AKI using KDIGO criteria, with systematic evaluation of renal function, hemolysis markers, and electrolyte balance. Hydration protocols and nephroprotective measures will be standardized, though specific strategies require clarification from the nephrology team. This research addresses critical gaps in evidence-based patient selection for invasive DVT treatment, particularly following the mixed results of the ATTRACT trial. The AI prediction model could enable personalized treatment decisions, potentially improving the risk-be