DSTransformer: A Dual-Scale Transformer Framework for Zero-Shot Blood Glucose Prediction via Domain Generalization

Diabetes is a global chronic metabolic disorder that requires precise blood glucose control to delay complications. Multi-step glucose prediction can provide early warnings of abnormal fluctuations and create a time window for proactive intervention. However, existing methods still face two major challenges: limited zero-shot generalization caused by patient heterogeneity, and difficulty in jointly capturing local fine-scale fluctuations and global long-term trends within a single-scale framework. To address these issues, this paper proposes DSTransformer, a dual-branch multi-resolution framework for multi-step blood glucose prediction. By combining parallel multi-scale feature extraction with layer-wise learnable routers, DSTransformer dynamically fuses local and global temporal information for more balanced glucose sequence modeling. It also introduces a Fishr-inspired patient-level gradient regularization strategy, treating each patient as an independent domain and aligning the loss landscape by matching batch-wise gradient variance across domains, which significantly enhances zero-shot generalization performance. Relying solely on univariate CGMS time series to keep clinical deployment costs low, the framework delivers strong results in patient-level zero-shot evaluations across three independent datasets (ShanghaiT2DM, OhioT1DM, and REPLACE-BG), predicting future blood glucose levels over 30-120-minute horizons. At the best performing 30-minute horizon, the average performance across the three datasets reached 13.335 mg/dL root mean square error (RMSE), 8.606 mg/dL mean absolute error (MAE), and 93.807% accuracy. Even for the more challenging 120 minute long-horizon prediction, the framework still achieved an average RMSE of 25.793 mg/dL, MAE of 16.995 mg/dL, and accuracy of 86.538% across the three datasets. These results outperform existing state-of-the-art (SOTA) methods across all key metrics, fully demonstrating the framework's excellent generalization capability and its potential to support future clinical decision-making after prospective clinical validation.