Ischemic stroke is one of the leading causes of disability and death, and early risk prediction is crucial for prevention and treatment. However, existing methods for predicting ischemic stroke have limitations, such as over-reliance on manually designed indicators and weights, use of only two-dimensional imaging data (e.g., ultrasound), and failure to integrate multiple patient examination results, leading to suboptimal prediction accuracy. Therefore, there is an urgent need for an intelligent prediction method that integrates multimodal information to enhance the clinical diagnosis of ischemic stroke. To address this, this project proposes a novel three-dimensional carotid artery computed tomography angiography (CTA) image segmentation model named CA-UNet, which fully automates the delineation of carotid arteries. Additionally, based on CA-UNet, an ischemic stroke risk prediction model is developed, leveraging patients' three-dimensional CTA images, electronic medical records, and medical history to predict stroke risk. A multi-scale loss function is designed to address the issue of detailed feature loss during downsampling, and transfer learning is employed with publicly available datasets to support the model. The results demonstrate that the proposed method achieves excellent performance in both carotid artery segmentation and ischemic stroke prediction tasks, with a Dice coefficient of 90.49% for carotid artery segmentation and an accuracy of 89.74% for ischemic stroke prediction. This method provides reliable diagnostic results for patients and healthcare professionals, offering significant clinical application value.