Dynamic Prediction of 3-Month Recurrence After First-Ever Ischemic Stroke Using a Two-Stream Attention-LSTM Model - Henan Province, China, 2025.

Introduction: We aimed to develop a dynamic model for predicting 3-month recurrence after first-ever ischemic stroke using multimodal longitudinal data. Methods: Patients with first-ever ischemic stroke admitted to a tertiary hospital in Zhengzhou, China between January 2023 and January 2025 were enrolled. Data on static baseline characteristics, discharge variables, and 1-month follow-up variables were collected. Imaging phenotypes were derived from diffusion-weighted imaging (DWI) and Fluid-Attenuated Inversion Recovery (FLAIR) using automated lesion segmentation and unsupervised clustering. The dataset was divided into training and validation sets (8∶2) using a fixed random seed. Seven models were trained and evaluated: logistic regression, random forest (RF), support vector machine (SVM), XGBoost, multi-layer perceptron (MLP), standard Long Short-Term Memory (LSTM), and two-stream attention-LSTM. External validation was conducted at three independent hospitals. The predictive performance was assessed using the area under the curve (AUC), accuracy, sensitivity, specificity, and Brier score. Results: Among 625 patients, recurrence occurred in 79 (12.64%) patients 3 months after discharge. In the validation set, the AUCs were ranked as follows: two-stream attention-LSTM (0.857); SVM (0.838); XGBoost (0.808); MLP (0.768); RF (0.761); standard LSTM (0.754); logistic regression with follow-up features (0.748); and discharge-only logistic regression (0.686). Two-stream attention LSTM identified key predictors, including dynamic changes in C-reactive protein, systolic blood pressure, and imaging phenotypes. External validation showed stable discrimination (pooled AUC=0.83) and good calibration. Conclusion: Two-stream attention-LSTM improved prediction of 3-month recurrence after first-ever ischemic stroke and may support early post-discharge risk stratification.