Systematic review and meta-analysis of machine learning models for deep vein thrombosis prediction in hospitalized adultsMachine learning models predict blood clots in hospitalized adults with high accuracy

ObjectivesTo systematically review machine learning–based prediction models for deep vein thrombosis (DVT) in hospitalized patients and to evaluate their methodological quality, predictive performance, and clinical applicability.MethodsThis systematic review was conducted in accordance with PRISMA and registered in PROSPERO (CRD420251069584). PubMed, Embase, Web of Science, and CINAHL were searched from November 1, 2015, to November 1, 2025. Studies developing, validating, or updating machine learning–based prediction models for DVT in hospitalized adults were included. Data were extracted on study characteristics, data sources, predictors, modeling methods, validation strategies, and model performance. Risk of bias and applicability were assessed using PROBAST. A meta-analysis of validation AUCs was performed in R, and heterogeneity was explored using the Cochrane Q test, I2 statistic, subgroup analysis, and sensitivity analysis.ResultsA total of 983 studies were initially identified through database searches. Following screening, 17 studies met the inclusion criteria, of which 10 were included in the meta-analysis. D-dimer, surgery, and age emerged as the most frequently reported predictors across models. Six studies were judged to be at high risk of bias, mainly because of limitations in the analysis domain, including inadequate reporting of missing-data handling, calibration, and validation procedures. The pooled AUC of the 10 validated models was 0.85 (95% CI: 0.81–0.90), indicating good overall discrimination; however, heterogeneity was substantial and remained high in subgroup and sensitivity analyses.ConclusionMachine learning–based models for predicting DVT in hospitalized patients show promising discriminative performance, but their clinical applicability remains limited by methodological weaknesses, poor reporting, substantial heterogeneity, and insufficient external validation. Future studies should prioritize multicenter external validation, transparent reporting, and head-to-head comparison with established clinical risk assessment tools.