Trend-weighted multi-resolution transformer for multi-parametric glucose prediction

Accurate glucose level prediction using continuous glucose monitoring data is crucial for effective diabetes management, enabling timely interventions to reduce hypo-and hyperglycaemic events. However, predicting future glucose levels remains a challenging due to the complex and variable nature of glucose dynamics. While artificial intelligence-based models have been widely applied to address these complexities, traditional approaches often lack the ability to capture the temporal dependencies. To address this limitation, we propose the Trend-Weighted Multi-Resolution Transformer (TW-MRT) model that effectively captures the complex temporal dependencies in glucose dynamics. TW-MRT introduces a trend-weighting mechanism to adaptively focus on relevant glucose trends, while a multi-resolution fusion layer processes data across varying time scales. This integrated layer allows the model to dynamically adjust glucose patterns, enhancing its predictive accuracy. The proposed model was evaluated on both the OhioT1DM and D1NAMO datasets, achieving root mean square error value of 11.37 mg/dL on OhioT1DM, and achieved an average RMSE of 11.30 mg/dL on D1NAMO for 30 PH. These results demonstrate the superior performance of the proposed TW-MRT model compared to existing state-of-the-art models. Furthermore, an ablation study confirms the contribution of each component within the model, demonstrating the importance of trend-weighting mechanism and multi-resolution fusion in improving predictive performance. Key contributions include: (1) trend-weighting and multi-resolution fusion enabling capture of short-, mid-, and long-term glucose dynamics, (2) cross-dataset validation demonstrating robust generalizability across diverse populations and sensor technologies, and (3) comprehensive clinical validation through surveillance error grid analysis, which confirms clinically reliable predictions across all patient scenarios.