<![CDATA[Introduction: Cerebral vascular malformations (CVMs) are the leading cause of spontaneous intracranial haemorrhage, yet their detection on CT/MR angiography is operator-dependent and existing machine-learning models derive almost exclusively from Caucasian or East Asian cohorts. We developed and internally validated a population-specific radiomics pipeline for CVM detection and rupture-risk stratification in a Southeast Asian population. Methods: In this retrospective diagnostic-and-prognostic cohort (STARD 2015; CLAIM) at a tertiary hospital in Palembang, Indonesia (2020–2025), 486 adults with diagnostic-quality CTA or TOF-MRA were analysed. After resampling and normalisation, 1,218 PyRadiomics features were reduced by LASSO and used to train Random Forest, SVM and XGBoost models (70/30 split). The reference standard was blinded consensus segmentation by two consultant neuroradiologists. Diagnostic accuracy (Wilson CI), AUC (DeLong), likelihood ratios, Cohen κ, McNemar test and a multivariable rupture model were computed. Results: CVM prevalence was 42.8% (208/486). XGBoost was the best detector (AUC 0.963 (95% CI 0.950–0.977); sensitivity 88.5 (95% CI 83.4–92.1)%; specificity 92.4 (95% CI 88.7–95.0)%; LR+ 11.71 (95% CI 7.74–17.72)), outperforming a single radiologist (ΔAUC 0.128, p<0.001; McNemar χ2=9.72, p=0.002 (discordant pairs b=73, c=39)). Inter-reader agreement was almost perfect (κ 0.845 (95% CI 0.797–0.893)). A radiomics-clinical model stratified rupture (AUC 0.846 (95% CI 0.792–0.901)), with lesion size (OR 4.26 (95% CI 2.60–6.98)) and hypertension (OR 2.46 (95% CI 1.18–5.14)) dominant and good calibration (χ2=14.87, df=8, p=0.062). Conclusion: A population-specific radiomics machine-learning pipeline achieved high diagnostic accuracy for CVM detection and clinically useful rupture-risk stratification, supporting operator-independent neurovascular triage in Southeast Asian settings. External validation is warranted.]]>
