Abstract Number: Esoc2026a2123 Predicting Malignant Cerebral Edema After Acute Ischemic Stroke Using Initial Diffusion-Weighted Mri

Background and aims To develop an automated framework integrating stroke lesion volume (SV) and cerebrospinal fluid volume (CSFV) for individualized prediction of malignant cerebral edema (MCE) after acute ischemic stroke (AIS) on diffusion-weighted imaging (DWI). Methods In this retrospective study, 530 consecutive patients with AIS (mean age, 68 years ± 14; 332 men), including 12 who developed MCE, were enrolled between January 2019 and October 2022. Stroke lesions were automatically segmented on initial DWI using apparent diffusion coefficient (ADC) threshold of 1.8 × 10-3 mm2/s to derive SV. CSFV was incorporated to represent the buffer zone. Two hierarchical feature sets were evaluated: one-dimensional (fixed thresholds: 82 mL, 145 mL), two-dimensional (SV + CSFV). Logistic regression (LR) and random forest (RF) classifiers were trained with patient-level separation and cross-validation. Model performance was assessed by accuracy, sensitivity, specificity, precision, and negative predictive value (NPV). Statistical comparisons were performed using Cochran’s Q testing with McNemar. P < 0.05 was considered statistically significant. Results LR and RF with two-dimensional features demonstrated comparable performance (accuracy, 0.995 vs. 0.994; sensitivity, 1.000 vs. 0.967; specificity, 0.995 vs. 0.943; precision, 0.803 vs. 0.375; NPV, 0.999 vs. 1.000). Both models significantly outperformed conventional volume-threshold methods (all p < .05 except sensitivity and specificity). Conclusions Automated DWI-based analysis integrating SV and CSFV enables highly accurate, individualized prediction of malignant cerebral edema after acute ischemic stroke and outperforms traditional SV-based criteria. Conflict of interest Cheng-Hsuan Juan, Chia-Hui Tsao, Ya-Hui Li, Chia-Ching Chang, Ming-Ting Tsai, Tung-Yang Lee, Cheng-En Juan, Hsiao-Wen Chung, Chun-Jung Juan: nothing to disclose