Hybrid CNN-ViT Modeling For Predicting Functional Outcome After Ischemic Stroke: A Retrospective Study.

Area of Science:

• Neurology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Acute ischemic stroke (AIS) is a leading cause of death and disability, necessitating accurate early prediction of functional recovery.
• Current methods for predicting 90-day functional outcomes (using modified Rankin Scale) based on early clinical and imaging data are challenging.
• Advanced predictive models are needed to guide clinical decisions and rehabilitation planning for AIS patients.

Purpose of the Study:

• To develop and evaluate a hybrid deep learning model integrating multiparametric MRI and clinical data for improved prediction of 90-day functional outcomes in AIS.
• To compare the performance of the hybrid model against single-modality clinical and imaging models.
• To assess the generalizability of the developed model on an independent external test cohort.

Main Methods:

• A retrospective cohort of 300 AIS patients underwent multiparametric MRI. Data were split into training-validation (n=250) and internal test (n=50) sets, with an external test set of 37 patients.
• A hybrid Convolutional Neural Network (CNN)-Vision Transformer (ViT) model was employed for extracting features from multiparametric MRI.
• Predictions from imaging and clinical data were integrated using stacked logistic regression for multimodal fusion.

Main Results:

• The hybrid multimodal fusion model achieved the highest internal test performance with an Area Under the Curve (AUC) of 0.885, sensitivity of 0.920, specificity of 0.760, and accuracy of 0.840.
• The best single clinical model (support vector machine) had an internal AUC of 0.878, while the imaging-only model achieved an AUC of 0.782.
• The multimodal fusion model demonstrated similar superior performance trends in the external test cohort, indicating good generalizability.

Conclusions:

• Stacked fusion of multiparametric MRI and clinical data using a hybrid CNN-ViT architecture significantly improves the prediction of 90-day functional outcomes in AIS patients.
• This multimodal approach offers a promising tool for enhancing prognostic accuracy in clinical practice.
• Further validation in larger, multicenter studies is recommended prior to widespread clinical implementation.