What question did this study set out to answer?

The central aim is to enhance predictive accuracy and stability in continuous blood glucose monitoring using a novel hybrid model.

March 17, 2026Scientific African0 citationsOpen Access

View Full Paper

Bidirectional recurrent-attention integrator network — A hybrid attention mechanism for accurate real-time continuous blood glucose monitoring in diabetes

NVNyirimana Jean Marie Vianney AMArchana Mathur DMDheeraj Reddy Mudireddy

Key Points

The central aim is to enhance predictive accuracy and stability in continuous blood glucose monitoring using a novel hybrid model.
Introduced Bidirectional Recurrent-Attention Integrator Network (BR-AIN) for glucose level predictions.
Evaluated model performance using RMSE, MAE, and MAPE across multiple forecast horizons.
Compared BR-AIN against baseline models (BiLSTM, LSTM, N-BEATS) using a multi-patient dataset.
Utilized Clarke Error Grid Analysis to assess clinical suitability of predictions.
BR-AIN achieved significantly lower RMSE and MAPE compared to baseline models at the 30-minute prediction horizon.
More than 95% of forecasts were categorized as clinically safe (Zones A and B).
Hybrid attention mechanism improved prediction stability and accuracy over longer forecasting periods.

Abstract

Continuous blood glucose monitoring plays a pivotal role in diabetes management, demanding precise and real-time predictive capabilities. This study introduces Bidirectional Recurrent-Attention Integrator Network (BR-AIN), a hybrid attention-based framework that leverages multi-head attention and hierarchical encoding mechanisms to enhance the accuracy and stability of glucose level predictions. The proposed model, BR-AIN, is evaluated on a multi-patient dataset using key performance metrics like Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) across multiple forecast horizons (15, 30, 45, and 60 min). A comparative analysis against benchmark models, including BiLSTM, plain LSTM, and standalone N-BEATS, demonstrates that the hybrid attention approach consistently achieves lower prediction errors and improved robustness, particularly in longer forecasting horizons. Visual analyses such as error heatmaps, boxplots of absolute error distributions, and rolling RMSE trends further support the model’s superior performance in capturing both short-term variability and long-term stability. We validated the clinical suitability of BR-AIN using Clarke Error Grid Analysis (CEGA) to categorize predictions into zones (A–E). The results show that BR-AIN produces clinically safe predictions, with more than 95% of forecasts falling within the clinically acceptable Zones A and B across all prediction horizons. The Wilcoxon signed-rank test on RMSE and MAPE at the 30-minute prediction horizon confirmed that BR-AIN achieves statistically significantly lower RMSE and MAPE than the baseline models, LSTM, BiLSTM, and N-BEATS. These findings highlight the potential of hybrid attention mechanisms as a dependable and expandable instrument for continuous real-time glucose monitoring in the treatment of diabetes. • Introduces BR-AIN, a hybrid recurrent-attention model for glucose monitoring. • Demonstrates superior accuracy across multiple forecast horizons. • Outperforms BiLSTM, LSTM, and N-BEATS baselines in robustness. • Supports diabetes management with real-time predictive stability.

KI fragen

Bookmark

View Full Paper

Cite This Study

Vianney et al. (Sun,) studied this question.

synapsesocial.com/papers/69b8ef52deb47d591b8c56ca https://doi.org/https://doi.org/10.1016/j.sciaf.2026.e03308

KI fragen

Bookmark

View Full Paper