Interpretable model using serum biomarkers and intra-/peri-aneurysmal CTA radiomics for discriminating ruptured from unruptured intracranial aneurysms: an external validation retrospective study.

The LRSLM fusion model discriminated ruptured from unruptured intracranial aneurysms with high performance in the external validation set.

• Among the 11 models, LRSLM achieved an external validation AUC of 0.9177 with good calibration.
• Accuracy was 0.8847, recall was 0.8171, and Brier score was low at 0.0999.
• LRSLM combined seven machine learning classifiers via logistic regression, compared with arithmetic mean, geometric mean, and weighted mean fusion strategies.

The LRSLM model provided greater clinical net benefit across a wide range of threshold probabilities.

• Decision curve analysis showed greater net benefit across threshold probabilities of 10-85%.
• Performance was evaluated in the external validation set (Center 2, N = 321).

Intra-aneurysmal and peri-aneurysmal radiomic features were strongly positively correlated.

• Spearman's r values were 0.62, 0.72, and 0.84 for the feature pairs analyzed (all p < 0.001).
• Peri-aneurysmal region was defined as a 1 mm region around the aneurysm.

Intra-/peri-aneurysmal radiomics were significantly positively correlated with the serum inflammation score (SIS).

• Spearman's r values were 0.32, 0.29, 0.32, and 0.34 (all p < 0.001).
• SIS was constructed from serum biomarkers, suggesting a link between imaging phenotype and systemic inflammation.

Restricted cubic spline analysis showed non-linear relationships between radiomic features and aneurysm rupture.

• Both intra-aneurysmal (IRSIAs) and peri-aneurysmal (PRSIAs) radiomic features were associated with increased rupture incidence.
• Non-linear p < 0.05 and Overall p < 0.05.

SHAP analysis identified key radiomic feature contributors to the fusion model and supported its interpretability.

• SHAP was used alongside correlation and RCS analyses to evaluate interpretability.
• The authors suggest it may help identify high-risk phenotypic features.

The study used a two-center retrospective design with external validation in 744 patients.

• Center 1 (N = 423) formed the training set; Center 2 (N = 321) formed the external validation set.
• Volumes of interest were delineated in intra-aneurysmal and peri-aneurysmal 1 mm regions on CTA.
• Seven machine learning classifiers were developed and combined using four fusion strategies, yielding 11 models in total.

The authors conclude that prospective validation is required before clinical predictive application.

• The study was retrospective.
• SHAP interpretability may assist in identifying high-risk phenotypic features.