Opportunities and Limitations of Artificial Intelligence in Phlebology: a Systematic Review with Meta-Analysis

Objective. To quantitatively evaluate the effectiveness of artificial intelligence (AI)-based technologies in diseases of the venous system. Material and methods. A systematic literature search was performed across the PubMed, Cochrane Library, and Google Scholar databases. Original studies utilizing AI models in venous pathology, including venous thromboembolism (VTE) and chronic venous disease (CVD), were included in the analysis. Study endpoints were the diagnostic and prognostic performance metrics, pooled using a random-effects model. Results. Of the 606 evaluated publications, 100 studies were included in the analysis addressing prediction (n=76) and diagnosis (n=24) of VTE (n=86) and CVD (n=14). For diagnostic models of VTE and CVD, the pooled accuracy, sensitivity, and specificity at internal validation were: 0.93 (95% CI: 0.87—0.96) and 0.98 (95% CI: 0.91—0.99); 0.89 (95% CI: 0.81—0.96) and 0.89 (95% CI: 0.80—0.95); 0.93 (95% CI: 0.87—0.98) and 0.94 (95% CI: 0.67—0.99), respectively. For predictive models of VTE, the pooled accuracy, sensitivity, specificity, precision, and AUC-ROC at internal validation were: 0.85 (95% CI: 0.83—0.88), 0.78 (95% CI: 0.72—0.85), 0.86 (95% CI: 0.80—0.92), 0.60 (95% CI: 0.45—0.74), and 0.86 (95% CI: 0.84—0.88), respectively. For predictive models of CVD, the pooled AUC was 0.88 (95% CI: 0.82—0.95). External validation was performed in only 23% of cases, and the obtained metrics were inferior to internal validation results. Conclusion. AI-based technologies demonstrate high diagnostic and moderate predictive performance in pathology of the venous system. However, further studies with external validation are required to assess implementation into clinical practice.