A fusion model of deep learning and conventional features based on computed tomography angiography of carotid plaque for predicting the risk of acute ischemic stroke

Background: Stroke is one of the leading causes of death and disability worldwide, with carotid atherosclerosis figuring prominently in its etiology. Traditional risk assessment based on stenotic degree provides only limited predictive power, while deep learning excels at mining latent features from medical images. This study aimed to develop and validate a fusion model integrating clinical risk factors, imaging features, and Swin Transformer-derived features from carotid plaque computed tomography angiography (CTA) for predicting ipsilateral acute ischemic stroke. Methods: We retrospectively analyzed 264 patients with carotid atherosclerosis who underwent head and neck CTA and magnetic resonance imaging (MRI). Patients were categorized into symptomatic and asymptomatic groups and randomly divided into training (n=184) and test (n=80) sets. Both DenseNet121 and Swin Transformer architectures were employed to extract deep learning features from carotid plaque CTA images. A conventional model was constructed from clinical risk factors and imaging features. A fusion model was developed through the integration of significant conventional features with Swin Transformer-derived features. Model performance was evaluated according to the area under the receiver operating characteristic curve (AUC), with the DeLong test being used for statistical comparisons. Decision curve analysis (DCA) and calibration curves were applied to assess clinical utility, while Shapley additive explanations (SHAP) and gradient-weighted class activation mapping (Grad-CAM) were used to enhance model interpretability. Results: In the test cohort, the conventional model had modest performance (AUC =0.691). Both deep learning models significantly outperformed the conventional model (P<0.05), with DenseNet121 achieving an AUC of 0.954 and Swin Transformer an AUC of 0.997. The fusion model also demonstrated excellent performance (AUC =0.989). Interpretability analyses revealed that the Swin Transformer model focused on biologically plausible regions corresponding to vulnerable plaque components. Conclusions: The integration of Swin Transformer-derived imaging features with conventional data significantly improves risk prediction for acute ischemic stroke as compared to conventional approaches. This multimodal framework provides a powerful and interpretable tool for enhanced risk stratification in patients with carotid atherosclerosis.