Predictive performance of CT-based artificial intelligence for predicting variceal bleeding in portal hypertension: a systematic review and meta-analysis.
Chao Zhu, Qi Liu, Wenhui Tao et al.
✦ AI-curated · Sources linked
In 30 seconds
This systematic review and meta-analysis assessed CT-based artificial intelligence models for predicting variceal bleeding in patients with portal hypertension. Ten studies involving 2,470 patients were included. Pooled sensitivity was 0.81 (95% CI: 0.73-0.87), specificity was 0.85 (95% CI: 0.75-0.91), and AUC was 0.88 (95% CI: 0.85-0.91). The authors describe CT-based AI as a promising adjunct for risk stratification, but note that internal validation and HBV-predominant cohorts limit confidence in generalisability.
Key findings
- Across included studies, pooled sensitivity was 0.81 (95% CI: 0.73-0.87) and specificity was 0.85 (95% CI: 0.75-0.91).
- The pooled AUC was 0.88 (95% CI: 0.85-0.91) for CT-based AI prediction of variceal bleeding.
- Unimodal image-only models had higher sensitivity than multimodal models (0.84 vs. 0.78).
- A Vision Transformer achieved an AUC of 0.98, but this result was limited to internal validation.
Why it matters
Variceal bleeding risk assessment may be supported by AI analysis of abdominal CT images already obtained in routine care. The pooled results suggest potential value for identifying lower-risk patients and informing endoscopy triage. However, the evidence does not establish that these models can replace established assessment or endoscopy, and performance may not generalise beyond the populations and validation settings studied.
What to keep in mind
The evidence relied on internal validation and HBV-predominant cohorts; the authors call for large-scale, international multicentre validation to establish generalisability.