Predicting malignant cerebral edema after acute ischemic stroke: a machine-learning model with multi-region radiomics

Area of Science:

• Radiology
• Artificial Intelligence
• Neurology

Background:

• Malignant cerebral edema (MCE) is a critical complication of acute ischemic stroke (AIS).
• MCE significantly increases the risk of poor outcomes or mortality in AIS patients.
• Accurate prediction of MCE is crucial for timely clinical intervention.

Purpose of the Study:

• To develop and validate a machine learning (ML)-based predictive model for MCE following AIS.
• To utilize radiomics features from non-contrast computed tomography (CT) images for MCE prediction.
• To compare the performance of a combined infarct, affected hemisphere, and whole brain (IWA) model against an infarct lesion (IL)-only model.

Main Methods:

• A total of 219 AIS patients were included from four centers, with data split into training, testing, and external validation cohorts.
• Radiomics features were extracted from the infarct lesion (IL), affected hemisphere (AH), and whole brain (WB).
• Seven ML algorithms were employed to develop an IL-only model and a combined IWA model, with performance assessed using the area under the curve (AUC).

Main Results:

• The combined IWA model demonstrated superior performance in predicting MCE risk compared to the IL-only model.
• The multilayer perceptron-based IWA model achieved a high AUC of 0.927, outperforming the IL model's AUC of 0.865.
• The developed IWA model showed statistically significant improvement (P<0.05) over the IL model.

Conclusions:

• A novel IWA radiomics model effectively predicts MCE risk in AIS patients.
• The IWA model significantly outperforms an IL-only model, offering improved predictive accuracy.
• This model is expected to enhance clinical decision-making and treatment guidance for AIS patients at risk of MCE.