Accurate glucose forecasting is essential for effective closed-loop insulin delivery in type 1 diabetes. A study published in Diabetologia compared the performance of personalized versus group-trained LSTM neural networks for glucose prediction.

Using data from 25 participants in the HUPA UCM dataset, researchers trained individualized and aggregated models and compared predictive accuracy. Personalized models, despite using much smaller datasets, achieved mean root mean squared error (RMSE) of 22.52 ± 6.38 mg/dL versus 20.50 ± 5.66 mg/dL for aggregated models. Clarke error grid Zone A accuracy was also similar (84.07% vs 85.09%).

Although a few individuals performed slightly better with group-trained models, overall differences were modest. These findings show that individualized neural networks can achieve clinically reliable glucose prediction with limited personal data. This approach may enable real-time adaptive learning on wearable devices or insulin pumps, improving precision, autonomy, and patient privacy in diabetes management.