Machine Learning and Deep Learning in Deep Vein Thrombosis: A Narrative Review of Clinical Applications for Diagnosis, Risk Prediction, and Outcome Management.

Abstract

Abstract: Machine learning (ML) and deep learning (DL) are increasingly applied across the clinical spectrum of deep vein thrombosis (DVT), from AI-guided ultrasound diagnosis to perioperative and inpatient risk prediction, but methodological progress has outpaced clinical translation. We aimed to narratively synthesize ML and DL applications in DVT across eight clinical domains and assess their performance and translational readiness. We performed a structured literature search of PubMed, Web of Science, and Scopus (January 2019-April 2026) using predefined disease and technology term blocks restricted to Title/Abstract fields. Of 492 records identified, 250 eligible publications were retained as the synthesis pool. Representative studies were purposively selected for detailed appraisal, prioritizing those with external validation, prospective design, or clear linkage to clinical workflows. AI applications span AI-guided ultrasound diagnosis, CT/MRI-based imaging, perioperative risk prediction, medical/ICU risk prediction, cancer-associated DVT, post-thrombotic syndrome and pulmonary embolism outcome prediction, and natural language processing (NLP)/large language model-based surveillance or decision support. Reported AUROC values are frequently high; however, external validation is uncommon, calibration metrics and decision-curve analysis are rarely reported, and prospective implementation studies are scarce. A randomized trial in medical inpatients demonstrated reduced hospital-acquired venous thromboembolism with AI-driven decision support, and AI-guided ultrasound pathways have shown near-expert diagnostic performance when used by non-expert operators. AI shows substantial promise for improving DVT diagnosis, risk stratification, and outcome prediction, but most models remain insufficiently validated for routine use. Prospective multicenter evaluation, standardized reporting of calibration and clinical utility, and structured oversight frameworks will be essential for safe deployment.