Abstract
Background: Type 1 Diabetes (T1D) management relies heavily on continuous glucose monitoring (CGM), which is susceptible to noise and missing values. Although simulators, such as the UVa-Padova model, provide clean datasets for training predictive machine learning (ML) models, they often fail to capture real-world physiological variations. Methods: This study evaluated three machine learning models—Linear Regression (LR), Long Short-Term Memory (LSTM), and eXtreme Gradient Boosting (XGBoost)—for glucose prediction using simulated and real-world CGM datasets. Predictive performance was assessed using statistical metrics and Clarke Error Grid Analysis (CEGA). Results: All evaluated models demonstrated mostly clinically safe predictions, although performance varied depending on the data source. The Linear Regression model consistently outperformed the LSTM model when trained with simulated datasets and achieved comparable predictive performance when trained and evaluated on real patient data. In contrast, XGBoost consistently demonstrated the worst performance across all tests. Conclusions: While simulators are valuable for baseline training, they do not fully reflect the complexities of real-world T1D management. The findings reveal that while simpler models suffice for simulated environments, capturing the physiological variability of real patients necessitates more complex architectures like LSTM, highlighting the critical need for real-world clinical validation in glucose prediction.