Key result
AI imaging models predict post-thrombectomy hemorrhagic transformation with high accuracy at ~0.89 pooled AUC.
Why the study?
Hemorrhagic transformation is a major complication after mechanical thrombectomy in acute ischemic stroke, and the overall predictive performance of artificial intelligence-driven imaging-based prediction models remains uncertain.
Do artificial intelligence-driven imaging-based prediction models accurately predict hemorrhagic transformation in acute ischemic stroke patients undergoing mechanical thrombectomy?
Meta-Analysis (n=3,541)
Yes
Do artificial intelligence-driven imaging-based prediction models accurately predict hemorrhagic transformation in acute ischemic stroke patients undergoing mechanical thrombectomy?
AI-driven imaging-based prediction models demonstrate favorable discrimination (pooled AUC 0.891, sensitivity 0.666, specificity 0.919) for predicting hemorrhagic transformation after mechanical thrombectomy in acute ischemic stroke, though high risk of bias and limited external validation preclude immediate clinical implementation.
Answers are grounded in this paper and cite their sources.
May aid post-thrombectomy risk stratification; leaves open routine adoption pending external validation and bias reduction.
Objective: Hemorrhagic transformation (HT) is a major complication after mechanical thrombectomy (MT) in acute ischemic stroke (AIS). Although artificial intelligence (AI)-driven imaging-based prediction models have shown promise for predicting HT, their overall predictive performance remains uncertain. This systematic review and meta-analysis evaluated their predictive performance. Methods: PubMed, Embase, Cochrane Library, Web of Science, CNKI, Wanfang, VIP, and CBM/SinoMed were searched from inception to July 2, 2026. Studies developing or validating AI-driven imaging-based prediction models for predicting HT after MT in AIS, including both pre- and post-procedural imaging predictors, were included. A random-effects model was used to pool the area under the curve (AUC), while a bivariate random-effects model was used to jointly estimate pooled sensitivity and specificity and construct the SROC curve. Subgroup analyses explored potential sources of heterogeneity. The bivariate random-effects model yielded a pooled sensitivity of 0.666 (95% CI: 0.578-0.744) and a pooled specificity of 0.919 (95% CI: 0.842-0.960). Exploratory subgroup analyses did not provide robust evidence supporting the superiority of any specific subgroup, and their interpretation was limited by small subgroup sizes and methodological heterogeneity. However, these findings should be interpreted cautiously because of the limited number of studies within individual subgroups. Conclusion: AI-driven imaging-based prediction models showed generally favorable discrimination in predicting HT after MT in AIS. However, current evidence does not establish the superiority of any specific model type. The predominance of retrospective studies, limited external validation, and methodological heterogeneity limit the generalizability of current findings. Large, multicenter prospective studies with rigorous external validation are needed before clinical implementation. Systematic review registration: The systematic review was registered with PROSPERO (CRD420261441463).
No takes yet. Share an insight, caveat, or question.
Hao et al. (2026) conducted a meta-analysis in Acute ischemic stroke treated with mechanical thrombectomy (n=3,541). AI-driven imaging-based prediction models (machine learning, deep learning, and radiomics) vs. Image-only vs. image-plus-clinical models, pre- vs. post-procedural imaging, or conventional risk scores/logistic regression was evaluated on Predictive performance for hemorrhagic transformation (HT), intracranial hemorrhage (ICH), or symptomatic intracranial hemorrhage (sICH) after mechanical thrombectomy (Pooled AUC 0.891, 95% CI 0.853-0.928, p=0.238 for heterogeneity). AI-driven imaging-based models predicted hemorrhagic transformation after stroke thrombectomy with a pooled AUC of 0.891, sensitivity of 0.666, and specificity of 0.919.
Synapse has enriched 10 related papers. Consider them for comparative context: