EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Accurate Prediction of Type 1 Diabetes Using a Novel Hybrid GRU-Transformer Model and Enhanced CGM Features

Loubna Mazgouti, Nacira Laamiri, Jaouher Ben Ali, Najiba El Amrani El Idrissi, Véronique Di Costanzo, Roomila Naeck, Jean-Mark Ginoux

Algorithms · 06.01.2026 · 19 (1) · 52

DOI: 10.3390/a19010052 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Algorithms, online erschienen am 06.01.2026. Verfasst von Loubna Mazgouti und 6 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 41 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate prediction of Blood Glucose (BG) levels is essential for effective diabetes management and the prevention of adverse glycemic events. This study introduces a novel designed hybrid Gated Recurrent Unit-Transformer (GRU-Transformer) model tailored to forecast BG levels at 15, 30, 45, and 60 min horizons using only Continuous Glucose Monitoring (CGM) data as input. The proposed approach integrates advanced CGM feature extraction step. The extracted features are statistically the mean, the median, the maximum, the entropy, the autocorrelation and the Detrended Fluctuation Analysis (DFA). In addition, in order to define more enhanced and specific features, the custom 3-points monotonicity score, the sinusoidal time encoding, and the workday/weekend binary features are proposed in this work. This approach enables the model to capture physiological dynamics and contextual temporal patterns of Type 1 Diabetes (T1D) with great accuracy. To thoroughly assess the performance of the proposed method, we relied on several well-established metrics, including Root Mean Squared Error (RMSE), Coefficient of Determination (R2), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Squared Percentage Error (RMSPE). Experimental results demonstrate that the proposed method achieves superior predictive accuracy for both short-term (15–30 min) and long-term (45–60 min) forecasting. Specifically, the model attained the lowest average RMSE values, with 4.00 mg/dL, 6.65 mg/dL, 7.96 mg/dL, and 8.91 mg/dL and yielding consistently high R2 scores for the respective prediction horizons. This new method distinguishes itself by continuously exceeding current prediction models, reinforcing its potential for real-time CGM and clinical decision support. Its high accuracy and adaptability make it a favorable tool for improving diabetes management and personalized glycemic control.

Zitieren

APA

Mazgouti, L., Laamiri, N., Ali, J. B., Idrissi, N. E. A. E., Costanzo, V. D., Naeck, R., & Ginoux, J. (2026). Accurate Prediction of Type 1 Diabetes Using a Novel Hybrid GRU-Transformer Model and Enhanced CGM Features. Algorithms, 19(1), 52. https://doi.org/10.3390/a19010052

Vancouver

Mazgouti L, Laamiri N, Ali JB, Idrissi NEAE, Costanzo VD, Naeck R, et al. Accurate Prediction of Type 1 Diabetes Using a Novel Hybrid GRU-Transformer Model and Enhanced CGM Features. Algorithms. 2026;19(1):52. doi:10.3390/a19010052

BibTeX
@article{Mazgouti2026,
  author = {Mazgouti, Loubna and Laamiri, Nacira and Ali, Jaouher Ben and Idrissi, Najiba El Amrani El and Costanzo, Véronique Di and Naeck, Roomila and Ginoux, Jean-Mark},
  title = {Accurate Prediction of Type 1 Diabetes Using a Novel Hybrid GRU-Transformer Model and Enhanced CGM Features},
  journal = {Algorithms},
  year = {2026},
  volume = {19},
  number = {1},
  pages = {52},
  doi = {10.3390/a19010052}
}
Kostenlos starten →