Akademik Veri Yönetim Sistemi
International Journal of Advanced Natural Sciences and Engineering Researches, cilt.9, sa.11, ss.16-22, 2025 (Hakemli Dergi)
Accurate short-term prediction of blood glucose levels is crucial for effective diabetes management, enabling timely interventions to prevent hypo- and hyperglycemic events and optimize insulin therapy. Glucose dynamics are inherently nonlinear and influenced by factors such as meal intake, insulin dosing, and individual metabolic variability, making accurate forecasting challenging. This study presents a Convolutional Long Short-Term Memory (ConvLSTM) deep learning framework for blood glucose prediction using simulated data from the SimGlucose environment, which provides physiologically realistic glucose–insulin dynamics across virtual patients in different age groups. The ConvLSTM model captures both local and temporal patterns by dividing input sequences into 12 subsequences of 12 time steps each. Convolutional layers extract features from each subsequence, while LSTM layers model sequential dependencies across the full sequence. The model was trained on continuous glucose monitoring (CGM) data and evaluated using Root Mean Square Error (RMSE) and Mean Squared Error (MSE). Across six test subjects, RMSE values ranged from 1.605 ± 0.562 mg/dL to 3.282 ± 1.074 mg/dL, demonstrating accurate and consistent predictions across children, adolescents, and adults. These findings indicate that the ConvLSTM architecture effectively learns complex glucose patterns and benefits from simulation-based datasets, which overcome the limitations of real-world CGM data such as scarcity, noise, and variability. By enabling controlled experimentation and reproducible evaluation, this approach supports the development of reliable, multi-step glucose forecasting models. Overall, the study highlights the potential of ConvLSTM networks combined with simulation-driven data for intelligent, data-driven diabetes management systems.