Malignant cerebral edema after endovascular thrombectomy: a multimodal prediction model based on post-thrombectomy cerebral hyperdensity and natural language processing.

BACKGROUND: Early prediction of malignant cerebral edema (MCE) following endovascular thrombectomy (EVT) is critical for guiding timely interventions. This study aimed to develop and validate a multimodal prediction, integrating non-contrast CT (NCCT) features and natural language processing (NLP)-encoded clinical data to predict MCE after EVT. METHODS: In this multi-center retrospective study, 373 patients treated with EVT were included, comprising internal (n = 287) and external (n = 86) cohorts. MCE was defined as a midline shift of ≥5 mm. Deep imaging features were extracted using a ResNet-101 model, the NCCT slice demonstrating the maximal extent of post-thrombectomy cerebral hyperdensity (PCHD). Concurrently, a pre-trained NLP model, BioClinicalBERT, was utilized to generate semantic embeddings from synthesized clinical narratives derived from standard admission variables. A multimodal fusion model integrating these features was subsequently evaluated against single-modality models and the diagnostic performance of human experts. RESULTS: In the independent external cohort, the multimodal fusion model achieved an area under the receiver operating characteristic curve (AUC) of 0.800 [95% confidence interval (CI): 0.700-0.901] and an accuracy of 80.2%, demonstrating superior performance compared to clinical-only (AUC = 0.654), ResNet-only (AUC = 0.707), and BERT-only (AUC = 0.560) models. SHapley Additive exPlanations (SHAP) analysis revealed NLP-derived semantic features as the principal predictors. Furthermore, AI assistance improved the diagnostic performance of senior neuroradiologists (AUC: 0.709-0.763; p < 0.05) and increased their specificity increased from (78.1% to 84.4%). CONCLUSION: A multimodal framework integrating targeted NCCT imaging features with NLP-encoded clinical data yields an accurate multimodal tool for early MCE prediction. This multimodal approach enhances human decision-making in emergency workflows.