This retrospective cohort study outlines a population-level approach to stroke prevention using real-world data and temporal AI. Stroke remains a global health challenge. Early identification of high-risk individuals can enable effective prevention. We developed a deep learning model to predict first-time stroke occurrence using temporal electronic health records (EHR) from Taiwan’s NHIRD (2003–2013). The model was trained on 16,805 incident stroke cases and 169,902 controls, utilizing structured binary matrices of ICD-9 and ATC codes across 3–24 months. A convolutional neural network (CNN) captured temporal patterns in diagnoses and prescriptions. Our model achieved an AUROC of 0.88 on the testing set using a 2-year observation window. To assess the impact of key predictors, we conducted a separate feature ablation analysis on the training data, which showed that removing the top-ranked medication feature (C08CA, a class of dihydropyridines) reduced the training AUROC from 0.91 to 0.85. These findings validate CNN’s ability to detect risk patterns in routine claims data. The model requires no additional tests and offers scalable risk stratification potential.