Personalized blood glucose prediction in type 1 diabetes using meta-learning with bidirectional long short term memory-transformer hybrid model
Kihoon Moon, Jaehong Kim, Seohyun Yoo, Jaehyuk Cho
Abstract
Personalized blood glucose (BG) prediction in Type 1 Diabetes (T1D) is challenged by significant inter-patient heterogeneity. To address this, we propose BiT-MAML, a hybrid model combining a Bidirectional LSTM-Transformer with Model-Agnostic Meta-Learning. We evaluated our model using a rigorous Leave-One-Patient-Out Cross-Validation (LOPO-CV) on the OhioT1DM dataset, ensuring a fair comparison against re-implemented LSTM and Edge-LSTM baselines. The results show our model achieved a mean RMSE of 24.89 mg/dL for the 30 min prediction horizon, marking a substantial improvement of 19.3% over the standard LSTM and 14.2% over the Edge-LSTM. Notably, our model also achieved the lowest standard deviation (±4.60 mg/dL), indicating more consistent and generalizable performance across the patient cohort. A key finding of our study is the confirmation of significant performance variability across individuals, a known clinical challenge. This was evident as our model's 30 min RMSE ranged from an excellent 19.64 mg/dL to a more challenging 30.57 mg/dL, reflecting the inherent difficulty of personalizing predictions rather than model instability. From a clinical safety perspective, Clarke Error Grid Analysis confirmed the model's robustness, with over 92% of predictions falling within the clinically acceptable Zones A and B. This study concludes that the development of effective personalized BG prediction requires not only advanced model architectures but also robust evaluation methods that transparently report the full spectrum of performance, providing a realistic pathway toward reliable clinical tools.