Multimodal ischemic stroke recurrence prediction model based on the capsule neural network and support vector...

Area of Science:

• Neurology
• Medical Imaging
• Machine Learning

Background:

• Ischemic stroke (IS) recurrence poses a significant clinical challenge.
• Current machine learning (ML) models for predicting IS recurrence, using single-modal data (biochemical or imaging), lack sufficient accuracy.

Purpose of the Study:

• To develop a high-performance, heterogeneous multimodal deep learning model for predicting ischemic stroke recurrence.
• To integrate magnetic resonance imaging (MRI) and biochemical test data for improved prediction accuracy.

Main Methods:

• A retrospective cohort study involving 634 IS patients with a 12-month follow-up for recurrence.
• Development of the ischemic stroke multi-group learning (ISGL) model, integrating capsule neural networks (for MRI T1/T2 signals) and a support vector machine (SVM) (for biochemical data).
• Comparison of the ISGL model against six classical ML and deep learning models.

Main Results:

• The ISGL model achieved high performance metrics: 95% accuracy, 96% specificity, 94% sensitivity, and 95% area under the curve (AUC).
• The ISGL model significantly outperformed other compared models, including the visual geometry group (VGG).
• Identified key biochemical risk factors: positive correlation with low-density lipoprotein, smoking, and heart disease history; negative correlation with total cholesterol, high-density lipoprotein, and diabetes.

Conclusions:

• The proposed ISGL model effectively predicts ischemic stroke recurrence by integrating multimodal data (MRI and biochemical tests).
• This multimodal approach offers superior predictive performance compared to single-modal ML models.
• The identified risk factors can guide early intervention strategies for high-risk patients, promoting secondary stroke prevention.