Multimodal deep learning for predicting unsuccessful recanalization in refractory large vessel occlusion

Area of Science:

• Neuroimaging and Artificial Intelligence
• Stroke Intervention and Outcomes
• Medical Data Fusion

Background:

• Acute ischemic stroke requires timely endovascular therapy (EVT) for effective treatment.
• Predicting EVT outcomes, especially in cases of refractory large vessel occlusion (rLVO), remains challenging.
• Integrating diverse patient data can enhance predictive accuracy for stroke interventions.

Purpose of the Study:

• To develop and validate a multi-modal deep learning model for predicting EVT outcomes in acute ischemic stroke patients.
• To assess the model's performance in identifying patients with refractory LVO (rLVO).
• To evaluate the contribution of integrated neuroimaging and clinical data compared to single-source models.

Main Methods:

• A retrospective analysis of 599 anterior circulation LVO patients undergoing EVT.
• Utilized non-contrast CT, CTA, CT perfusion, and clinical data.
• Employed convolutional neural networks for imaging feature extraction and a DAFT module for data fusion.

Main Results:

• The multi-modal model achieved an AUC of 0.70 ± 0.02 and F1 score of 0.39 ± 0.02 in predicting rLVO.
• This significantly outperformed models using only imaging data (AUC 0.53 ± 0.02, F1 0.19 ± 0.05).
• Explainability methods identified key clinical variables and image regions influencing predictions.

Conclusions:

• Combining vascular segmentation, clinical variables, and imaging data enhances prediction performance for EVT outcomes.
• This approach can provide early alerts for procedural complexity.
• It holds potential for guiding tailored and timely intervention strategies in EVT workflows.