Predictive performance of CT-based artificial intelligence for predicting variceal bleeding in portal hypertension: a systematic review and meta-analysis.

OBJECTIVES

To systematically evaluate the predictive accuracy of computed tomography (CT)-based artificial intelligence (AI) for predicting variceal bleeding (VB) in patients with portal hypertension, and to assess their potential utility as an opportunistic screening tool alongside Baveno VII criteria.

METHODS

We searched PubMed, Embase, Web of Science, and Cochrane until December 16, 2025, for studies applying radiomics or machine learning algorithms to abdominal CT images for VB prediction. Quality was assessed using PROBAST + AI. A bivariate random-effects model was employed to calculate pooled sensitivity, specificity, and the area under the curve (AUC).

RESULTS

Ten studies encompassing 2,470 patients were included. CT-based AI models demonstrated promising predictive performance with a pooled sensitivity of 0.81 (95% Confidence Interval [CI]: 0.73-0.87), specificity of 0.85 (95% CI: 0.75-0.91), and an AUC of 0.88 (95% CI: 0.85-0.91). Unimodal image-only models achieved higher sensitivity than multimodal models (0.84 vs. 0.78). While a Vision Transformer architecture achieved the highest accuracy (AUC 0.98), it was limited to internal validation. Fagan's nomogram indicated a negative likelihood ratio of 0.23, reducing an assumed post-test probability of bleeding from 20% to 5%.

CONCLUSION

CT-based AI models exhibit high predictive efficacy and offer a promising non-invasive "gatekeeper" strategy for risk stratification. By leveraging routine imaging, these models may reduce unnecessary endoscopies for low-risk patients. However, given the reliance on internal validation and HBV-predominant cohorts, results should be interpreted as valuable adjunctive evidence rather than a standalone replacement. Large-scale, international multi-center validation is required to confirm generalizability before clinical implementation.