CT hyperdensity and NLP model predicts malignant edema after stroke thrombectomy

A multimodal model combining post-thrombectomy CT hyperdensity features and NLP-clinical embeddings predicted malignant cerebral edema (midline shift ≥5 mm) with AUC 0.800 (95% CI 0.700-0.901) on external validation, improving senior neuroradiologists’ specificity from 78.1% to…

• The fusion model outperformed single-modality models: clinical-only AUC 0.654, imaging-only AUC 0.707, NLP-only AUC 0.560.
• AI assistance improved senior neuroradiologists’ AUC from 0.709 to 0.763 (p<0.05) and specificity from 78.1% to 84.4%.
• SHAP analysis identified NLP-derived semantic features as the top predictors.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.