Background: Type 1 Diabetes Mellitus (T1DM) is a chronic condition where insulin deficiency and frequent hypoglycemia remain barriers to glycemic control. Clinical prediction tools like threshold detection underperform due to physiologic variability, and high false alarm rates causing alarm fatigue. Artificial intelligence (AI) advancements offer significant potential for improved hypoglycemia prediction.

Methods: The T1DiabetesGranada dataset includes continuous glucose measurements, biochemical parameters, and demographic information from 736 adults with T1DM, covering 257,780 patient-days. A Temporal Fusion Transformer with attention mechanisms and temporal convolutional layers was trained to map multimodal signal features to short-term glucose trajectories. Machine learning optimization and cross validation enhanced predictive accuracy. Model performance was assessed on unseen subjects and compared against glucose-only neural networks and clinical threshold methods for 30-minute hypoglycemia forecasting.

Results: The TFT model achieved a 30-minute hypoglycemia prediction accuracy of 91.2% (AUC = 0.94), outperforming existing glucose-only methods (AUC = 0.81, p < 0.001) while reducing false alarms by 42%. Key predictive features included temporal patterns and cross-signal interactions captured by the attention mechanism. Compared to conventional glucose-only deep learning models, the TFT showed greater sensitivity to early physiologic changes preceding hypoglycemia. Against clinical threshold-based detection models, it achieved substantially fewer false positives and better detection during nocturnal and rapid glycemic fluctuations. Performance remained consistent across a clinically relevant range of glycemic control (HbA1c 6.5—8.5%). Real-time inference on a mobile device averaged 0.12 seconds per prediction, supporting continuous wearable deployment.

Discussion/Conclusion: The attention-augmented Temporal Fusion Transformer significantly improves short-horizon hypoglycemia prediction compared to existing methods. By drastically decreasing false alarm rates while maintaining high prediction accuracy, our method offers a scalable solution for anticipatory glucose management in T1DM. These results highlight the potential of transformer-based AI models to provide clinically reliable, real-time decision support, reducing hypoglycemic burden while improving safety and reliability.

@article{ma2026attention,
  title={Attention-Augmented Temporal Fusion Transformer for Short-Horizon Hypoglycemia Prediction in Type 1 Diabetes},
  author={Ma, Andrew and Chennareddy, Srihith},
  journal={Endocrine Practice},
  volume={32},
  pages={S336--S347},
  year={2026},
  month={March},
  publisher={Elsevier},
  doi={10.1016/j.eprac.2026.03.018},
  url={https://doi.org/10.1016/j.eprac.2026.03.018}
}