Abstract:
Diabetes is a critical global health issue affecting millions worldwide, with cases steadily increasing. Our research focuses on blood glucose levels prediction using a CRNN (Convolutional Recurrent Neural Network) model applied to a newly accessible and extensive dataset, "HUPA-UCM Diabetes," aiming to identify hyperglycemia and hypoglycemia events to enhance diabetes management accurately. This hybrid model integrates CNN and LSTM layers to effectively capture spatial and temporal dependencies in the data, achieving superior accuracy. It utilizes variables such as time, glucose, calories, heart rate, steps, basal rate, bolus volume delivered, and carb input. Our model demonstrated an RMSE value of 3.20 on the testing set, outperforming our state-of-the-art. Furthermore, real-time processing, implemented at five-minute intervals, ensures immediate responses to blood sugar variations with an RMSE of 4.63, improving patient outcomes.
