Diagnostic performance of machine learning models versus established risk stratification for intracranial aneurysm rupture: a systematic review and bivariate meta-analysis.
Shaan Patel, Shiva A Nischal, Yi-Hein Chai et al.
✦ AI-curated · Sources linked
In 30 seconds
This systematic review and meta-analysis evaluated the diagnostic performance of machine learning (ML) models for predicting intracranial aneurysm rupture compared to established risk stratification tools. Analyzing 62 cohorts with 29,709 patients, ML models showed pooled sensitivity of 0.81 and specificity of 0.83, with an area under the curve (AUC) of 0.878, outperforming the PHASES score (AUC 0.667).
Key findings
- Pooled sensitivity and specificity for ML models were 0.81 and 0.83, respectively.
- ML models achieved an AUC of 0.878, surpassing the PHASES score's AUC of 0.667.
- In external validation, ML sensitivity was 0.82, but specificity dropped to 0.66.
- Incorporating haemodynamic or radiomic features improved discrimination compared to morphology alone.
Why it matters
The findings suggest that machine learning models may enhance the prediction of intracranial aneurysm rupture compared to traditional risk scores, potentially leading to better patient management. However, the reduced specificity in external validation raises concerns about their immediate clinical applicability.
What to keep in mind
Evidence of small-study effects and unclear risk of bias ratings were noted, which may affect the reliability of the findings.