Key result
Most AI models for VTE carry high bias risk, with ~90% failing rigorous quality standards.
Why the study?
Artificial intelligence has potential to improve VTE care, but existing studies show high risk of bias and lack prospective validation and calibration reporting.
Systematic Review (n=23)
While AI shows promise for enhancing VTE prediction and detection, the vast majority of current models suffer from a high risk of bias and lack rigorous external validation, highlighting the need for standardized evaluation before clinical deployment.
Clinicians should not yet use AI VTE models in practice; leaves open safe deployment until external validation and bias reduction become standard.
Artificial intelligence (AI) has the potential to support personalized, multidisciplinary, data-driven care for venous thromboembolism (VTE) prevention, detection, imaging, management, and follow-up. We reviewed 23 studies: 20 prediction or detection studies were assessed with PROBAST+AI, while 3 evaluated clinical-impact interventions. Of these, 18 of 20 studies (90.0%) were at high risk of bias. Only 7 studies (35%) showed meaningful external or prospective validation, though 5/7 remained high risk of bias. Calibration was not reported in 14 studies (70.0%). Discrimination scores ranged from modest to high, but performance varied across cohorts and was often limited by retrospective design, internal validation, class imbalance, low event counts, and overfitting. Currently, agentic AI remains conceptual and needs human oversight. Future research should emphasize prospective multicenter evaluation, calibration, transparent reporting, regulatory adherence, and standardized bias auditing before AI is used routinely in clinical practice.
No takes yet. Share an insight, caveat, or question.
Moreno et al. (2026) conducted a systematic review in Venous thromboembolism (n=23). Artificial intelligence models vs. Standard care or traditional risk scores was evaluated on High risk of bias assessed with PROBAST+AI in prediction or detection studies. A review of artificial intelligence in venous thromboembolism care found that 90% of prediction or detection models were at high risk of bias, with only 35% showing meaningful external validation.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: