Generalizing CGM Sensor-Based Glucose Prediction across Age Cohorts Using LSTM Models: An In Silico Study with the UVA/Padova T1DMS Simulator

Accurate glucose prediction is necessary in enhancing insulin therapy and preventing harmful blood-sugar spikes in individuals with type 1 diabetes (T1D). Although deep learning models seem promising for glucose forecasting, it is unclear how they perform across various age groups that exhibit different metabolism profiles. This paper compares the performance of the long short-term memory (LSTM) models across age groups using simulated data from the UVA/Padova T1D Metabolic Simulator (T1DMS). Cohort-specific models achieved high withincohort performance (MAE < 2 mg/dL, r > 0.99, R² > 0.99), indicating precise modeling of glucose–insulin dynamics within each group. Nevertheless, predictive accuracy decreased when models were applied across cohorts, especially when LSTM networks trained on adults were tested on younger groups, demonstrating physiological variability between ages in insulin sensitivity and glucose kinetics. Models trained on younger groups performed better on older populations, suggesting that a broader range of metabolic variation underlies increased adaptability. This study is the first to use the FDA-approved UVA/Padova T1DMS simulator to systematically assess age-dependent generalization in LSTM-based glucose prediction, offering a unique reproducible framework for developing adaptive e-health and closedloop insulin systems. Incorporating age-relevant physiological heterogeneity and adaptive modeling paradigms could help develop stronger, patient-specific glucose forecasting systems for safer and more effective diabetes management.