SparGlu: A sparse‐data evaluation framework for gated recurrent unit‐based blood glucose prediction across varying monitoring intervals

Abstract Blood glucose prediction can support proactive diabetes management, but most existing models are developed using high‐frequency continuous glucose monitoring (CGM) data. Their reliability under sparse monitoring conditions, which are common in resource‐limited settings, remains insufficiently characterized. This study proposes SparGlu, a sparse‐data evaluation framework using GRU‐based deep learning model to forecast glucose dynamics across progressively reduced monitoring frequencies. Using the Ohio type 1 diabetes mellitus dataset, CGM, insulin, and meal‐related variables were extracted, temporally aligned, and evaluated at the original 5‐min resolution and after down sampling to 30 min, 1 h, 90 min, 2 h, 3 h, and 4 h. For the original 5‐min data, the prediction horizon (PH) was 30 min; for down‐sampled data, one‐step‐ahead prediction was used, so the PH equaled the sampling interval. The model achieved the strongest performance at 5‐ and 30‐min intervals, with Mean Absolute Error values of approximately 13–15 mg/dL, R 2 above 0.80, and more than 95% of predictions in Clarke zones A + B. At the 1‐h interval, performance declined, with Clarke A + B decreasing to approximately 90%. Performance degraded further at intervals of 90 min and longer. These findings suggest that sparse‐data glucose prediction is feasible up to moderate levels of sparsity, but clinical deployment requires cautious interpretation and clinical validation.