AI imaging models predict stroke bleeding after thrombectomy, review finds

Artificial intelligence models that read brain scans may help predict one of the most feared complications of stroke treatment: bleeding into the damaged brain tissue after a clot is mechanically removed. A new systematic review and meta-analysis, published in Frontiers in Neurology, pooled data from eleven studies covering 3,541 patients and found that AI-driven imaging-based prediction models achieved a pooled area under the curve of 0.891, with a 95 percent confidence interval of 0.853 to 0.928. That figure describes discrimination, the ability of a model to separate patients who later develop hemorrhagic transformation from those who do not, and it was accompanied by low between-study heterogeneity, with an I-squared of 29.0 percent and a p value of 0.238.

The complication in question is hemorrhagic transformation, a spectrum of bleeding that ranges from asymptomatic hemorrhagic infarction to symptomatic parenchymal hematoma. According to the review, it occurs in roughly 30 to 50 percent of patients following mechanical thrombectomy, the catheter-based procedure used to pull large clots out of brain arteries in acute ischemic stroke. The most severe form, symptomatic intracranial hemorrhage, has been reported in 5 to 10 percent of patients and is strongly associated with poor functional outcomes and increased mortality. Because therapeutic options are limited once bleeding begins, the authors argue that early and accurate risk prediction could support risk stratification and closer neurological monitoring around the time of the procedure.

Mechanical thrombectomy itself is firmly established. Multiple randomized controlled trials have demonstrated its efficacy in improving functional outcomes for acute ischemic stroke caused by large vessel occlusion, and international guidelines strongly recommend it for eligible patients with anterior circulation occlusion within 24 hours of symptom onset. The problem is that restoring blood flow to tissue that has been starved of oxygen carries an inherent risk: reperfusion can damage the blood-brain barrier, trigger inflammatory cascades, and set the stage for vessels to leak. The pathophysiology of hemorrhagic transformation is described as complex and multifactorial, involving blood-brain barrier disruption, reperfusion injury, endothelial dysfunction, and post-recanalization inflammation.

Traditional approaches to predicting this bleeding have relied on clinical risk scores such as TAG, ASIAN, and SEDAN, on individual markers like the ASPECTS score and admission glucose, or on conventional logistic regression models. The review notes that the performance of these approaches is often modest, particularly in external validation cohorts, and that conventional methods may struggle to capture complex nonlinear interactions among multiple predictors. Artificial intelligence techniques, including machine learning, deep learning, and radiomics, offer a different route: they can automatically extract high-dimensional quantitative features from medical images and recognize nonlinear patterns, potentially reducing reliance on subjective visual interpretation by radiologists and clinicians.

The meta-analysis, registered with PROSPERO under CRD420261441463, searched eight databases, including PubMed, Embase, the Cochrane Library, Web of Science, and four Chinese databases, from inception to July 2, 2026. The eleven included studies were published between 2023 and 2026 and came from China, the United States, Korea, and Brazil. All were cohort studies, and the majority were retrospective. Four were multicenter, and only three included independent external validation. Sample sizes ranged from 105 to 922 patients, with a median of 288, and the proportion of hemorrhagic outcomes within the study cohorts ranged from 7.7 to 70.1 percent, with a median of 39.8 percent.

The models themselves varied considerably. Imaging modalities included non-contrast computed tomography, CT angiography, CT perfusion, multiparametric MRI, dual-energy CT, and digital subtraction angiography, acquired either before the procedure or after it. Predictor sets ranged from image-only inputs to combinations of imaging with clinical variables. Algorithms included convolutional neural networks, radiomics pipelines paired with LASSO variable selection and logistic regression, support vector machines, XGBoost gradient boosting, and a TabPFN transformer model. Outcome definitions also differed across studies, spanning radiological hemorrhagic transformation, intracranial hemorrhage, symptomatic intracranial hemorrhage, and, in one study, a broad composite that included contrast extravasation.

Beyond the pooled AUC, the authors ran a bivariate random-effects model on the eight studies that reported usable paired sensitivity and specificity data. That analysis yielded a pooled sensitivity of 0.666, with a 95 percent confidence interval of 0.578 to 0.744, and a pooled specificity of 0.919, with a confidence interval of 0.842 to 0.960. In plain terms, the models as a group were better at ruling out bleeding than at catching every case: they rarely flagged patients who did not go on to bleed, but they missed roughly a third of those who did. Three studies were excluded from this bivariate analysis for technical reasons, including one whose reported sensitivity and specificity could not be reconciled with a unique contingency table and another whose performance was averaged across repeated random train-test splits.

Exploratory subgroup analyses probed whether performance varied by imaging timing, predictor type, validation type, outcome definition, or country. One nominal difference emerged: post-procedural imaging models showed a pooled AUC of 0.927, compared with 0.861 for pre-procedural models, with a subgroup difference p value of 0.048. The authors urge strong caution here. Because post-procedural images are acquired close in time to the hemorrhagic outcome assessment, some imaging features may already reflect early hemorrhagic changes or contrast extravasation, meaning the apparently higher discrimination could partly represent early detection of an evolving bleed rather than genuine prediction of a future event. The comparison also involved only five studies per arm and was one of five subgroup tests performed without multiplicity adjustment.

No statistically significant subgroup differences were found for the other comparisons. Image-plus-clinical models showed a numerically higher pooled AUC than image-only models, 0.907 versus 0.860, but the difference was not significant at p equal to 0.151, so the current evidence does not establish a clear advantage from adding clinical variables. Externally validated models scored 0.915 versus 0.872 for internally validated ones, again without statistical significance. Head-to-head comparisons inside individual studies were mixed: one XGBoost model beat logistic regression and several clinical scores in a United States cohort, a TabPFN model outperformed the ASIAN and TAG scores in external validation but did not significantly beat logistic regression, and another study found no significant differences in DeLong comparisons. The review concludes that the available evidence does not demonstrate consistent superiority of AI models over conventional statistical approaches.

The quality assessment highlights why caution is warranted. Using the PROBAST+AI tool, the reviewers rated ten of the eleven studies as having a high overall risk of bias and one as unclear, with the analysis domain the dominant problem: nine studies were rated high risk there, driven by limited outcome events relative to model complexity, model selection using evaluation datasets, small validation samples, and incomplete reporting of calibration and uncertainty. Calibration and decision-curve analyses were reported in only five studies each, with methods too heterogeneous to pool. The authors state that discrimination alone does not establish clinical utility and that the current evidence is insufficient to support changes in treatment decisions based on these models. They call for large, multicenter, prospective studies with rigorous external validation, calibration assessment, and evaluation of clinical utility before AI-driven hemorrhagic transformation prediction can move from research prototypes into routine stroke care.

Subject of Research: AI-based imaging prediction of hemorrhagic transformation after mechanical thrombectomy in acute ischemic stroke

Article Title: Artificial intelligence-driven imaging-based prediction models for hemorrhagic transformation after mechanical thrombectomy in acute ischemic stroke: a systematic review and meta-analysis

Article References: Hao, P., Wang, S., & Xin, R. (2026). Artificial intelligence-driven imaging-based prediction models for hemorrhagic transformation after mechanical thrombectomy in acute ischemic stroke: a systematic review and meta-analysis. Frontiers in Neurology, 17, Article 1960029. https://doi.org/10.3389/fneur.2026.1960029

Image Credits: AI Generated

DOI: 10.3389/fneur.2026.1960029

Keywords: acute ischemic stroke, mechanical thrombectomy, hemorrhagic transformation, artificial intelligence, machine learning, radiomics, medical imaging, meta-analysis, predictive models, symptomatic intracranial hemorrhage, external validation, risk of bias