Accelerating Prediction of Malignant Cerebral Edema After Ischemic Stroke with Automated Image Analysis and...

Area of Science:

• Neurology
• Radiology
• Artificial Intelligence in Medicine

Background:

• Malignant cerebral edema is a severe stroke complication requiring timely intervention.
• Current prediction methods often rely on advanced imaging and have limited accuracy.
• Early identification of patients at risk is crucial for preventing herniation and death.

Purpose of the Study:

• To evaluate the efficacy of neural networks using routine computed tomography (CT) data for predicting malignant cerebral edema.
• To compare the performance of deep learning models against traditional regression models and the EDEMA score.

Main Methods:

• Retrospective analysis of volumetric data (CSF volumes, hemispheric CSF ratio) from CT scans of a large stroke cohort.
• Training of fully connected and long short-term memory (LSTM) neural networks using clinical and imaging data.
• Performance evaluation using cross-validation and precision-recall curves.

Main Results:

• The LSTM model achieved 100% recall and 87% precision in predicting malignant edema, significantly outperforming regression models and a fully connected network.
• Key predictive features included the hemispheric CSF ratio and National Institutes of Health Stroke Scale (NIHSS) score at 24 hours.
• The LSTM model demonstrated superior performance (AUPRC 0.97) compared to regression (AUPRC 0.74) and fully connected models (AUPRC 0.71).

Conclusions:

• A deep learning framework, specifically an LSTM neural network, can accurately identify malignant cerebral edema using routine CT scans within 24 hours post-stroke.
• This approach shows promise for assisting in the selection of patients for hemicraniectomy, potentially improving outcomes.
• Prospective validation is required, but the study provides proof of principle for AI-assisted stroke management.