Stroke-bleeding AI models looked accurate—why is clinical use still premature?
A meta-analysis of 11 cohort studies and 3,541 patients found pooled AUC 0.891 for imaging-based models after thrombectomy. High risk of bias, limited external validation and mixed comparators leave clinical benefit unproven.
The short version
- What happened
- The review included 11 cohort studies with 3,541 patients; ten studies contributed eligible AUC estimates and evaluation cohorts ranged from 18 to 261 people.
- Who it affects
- A reliable risk tool could help stroke teams target monitoring after thrombectomy, but missing a high-risk patient can be consequential and pooled sensitivity was about two-thirds.
- What remains unproven
- Most included studies were retrospective, and only three reported independent external validation.
Keep in mind
Most included studies were retrospective, and only three reported independent external validation.
The review judged most studies at high risk of bias, especially because of analysis choices, small samples and limited event counts.
Outcome definitions, imaging timing, model workflows and comparators varied across studies.
Post-procedure images may capture early bleeding rather than predict a future event, creating predictor–outcome temporal overlap.
The meta-analysis assessed discrimination, not whether model-guided care improves disability, survival or quality of life.
Source analysis · 1 linked source. Primary analysed source date: 20 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The pooled models ranked risk well, with uneven sensitivity
Researchers searched eight English- and Chinese-language databases through 2 July 2026 for imaging-based artificial-intelligence models predicting hemorrhagic transformation after mechanical thrombectomy. Eleven cohort studies with 3,541 patients met the review criteria; ten supplied an eligible area-under-the-curve estimate. Evaluation cohorts were small, ranging from 18 to 261 people.
The pooled AUC was 0.891, with a 95% confidence interval of 0.853–0.928 and I² of 29.0%. Pooled sensitivity was 0.666 and specificity was 0.919. In plain language, the reviewed models tended to rank higher-risk patients well and produced fewer false positives than false negatives. But the studies used different imaging, outcome definitions, time points and modelling workflows, and direct comparisons with conventional methods did not establish consistent superiority.[1]
Bias and timing limit what the summary number means
Most studies were retrospective, only three reported independent external validation, and the review judged most studies at high risk of bias—particularly in the analysis domain. Small validation sets, many candidate predictors relative to outcome events and incomplete reporting can make performance look better than it will be in new patients. With only 11 studies, the reviewers did not formally test publication bias.
There is also a clinical-timing problem: some post-procedural images may already contain early signs of bleeding, so a model can partly detect an event that has begun rather than predict a genuinely future complication. No pooled result shows that using a model changes blood-pressure management, antithrombotic treatment, intensive monitoring, disability or survival. Clinicians therefore need prospective, locked-model validation and a trial of model-guided decisions before these scores influence care.[1]
The distinction that matters
What the meta-analysis found—and what it did not
The headline performance is a pooled discrimination result from small observational cohorts, not a clinical-effect estimate.
Evidence base
11 cohorts; 3,541 patients
10 eligible AUC estimates
Performance
AUC 0.891
Sensitivity 0.666; specificity 0.919
Transport
3 externally validated studies
Clinical benefit not tested