Deep learning on pre-procedural computed tomography and clinical data predicts outcome following stroke thrombectomy

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Imaging

Background:

• Pre-treatment prognostication is crucial for ischemic stroke patients undergoing endovascular thrombectomy (EVT).
• Deep learning (DL) models offer potential for improved prediction using clinical and imaging data.

Purpose of the Study:

• To evaluate the efficacy of DL models in predicting 3-month functional outcomes in EVT patients.
• To compare DL model performance against classical machine learning and existing prognostic tools.

Main Methods:

• DL models were trained and tested on baseline clinical and imaging (CT head, CT angiography) data.
• Classical models (logistic regression, random forest) and the MR PREDICTS tool were used for comparison.
• External validation was performed on an independent dataset.

Main Results:

• A deep learning model using CT and clinical data achieved strong discriminative ability (AUC 0.811).
• Logistic regression using clinical data alone showed comparable performance (AUC 0.817).
• Both models outperformed other DL approaches and the MR PREDICTS tool.

Conclusions:

• Deep learning demonstrates comparable discriminative performance to logistic regression for predicting functional independence post-EVT.
• Future research should explore the impact of procedural and post-procedural data on model accuracy.