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

Objective To develop an interpretable model that integrates diverse machine learning algorithms, fusion strategies, intra−/perianeurysmal CTA radiomics, and serum biomarkers to differentiate ruptured from unruptured intracranial aneurysms (IAs), and to investigate the underlying biological features associated with rupture status. Methods Totally, 744 IAs patients who underwent CTA were retrospectively enrolled. Data from Center 1 ( N = 423) served as the training set, and data from Center 2 ( N = 321) served as the external validation set. Volumes of interest were delineated in intra-aneurysmal and peri-aneurysmal 1 mm regions. The serum inflammation score (SIS) was constructed using serum biomarkers. Seven machine learning classifiers were employed to develop models, and the optimal fusion model was developed by combining all classifiers using arithmetic mean, geometric mean, weighted mean, and logistic regression (LRSLM). Model performance was assessed by the area under the curve (AUC) and decision curve analysis (DCA), while interpretability was evaluated via SHAP, correlation, and restricted cubic spline (RCS) analyses. Results Among the 11 models, LRSLM exhibited robust performance in the external validation set, achieving an AUC of 0.9177 with good calibration. It also demonstrated high accuracy (0.8847), recall (0.8171), and a low Brier score (0.0999), alongside greater clinical net benefit across a wide range of threshold probabilities (10–85%). Correlation analyses revealed strong positive associations between intra-aneurysmal and peri-aneurysmal radiomic features (Spearman’s r = 0.62, 0.72, and 0.84, respectively; all p < 0.001), as well as significant positive correlations between the intra−/perianeurysmal radiomics and SIS (Spearman’s r = 0.32, 0.29, 0.32, and 0.34; all p < 0.001). RCS analysis further demonstrated a significant non-linear relationship between the intra-aneurysmal (IRSIAs) and peri-aneurysmal (PRSIAs) radiomic features and an increased incidence of IAs rupture (Non-linear p < 0.05; Overall p < 0.05). SHAP analysis highlighted key radiomics features contributors to the fusion model and supported model interpretability. Conclusion The LRSLM model achieves robust discriminative performance for differentiating ruptured from unruptured IAs. SHAP analysis enhances the interpretability of the model and may assist in identifying high-risk phenotypic features, although prospective validation is required before predictive clinical application can be considered.