Long-term blood glucose prediction using deep learning-based noise reduction

Abstract

The Artificial Pancreas System (APS) is a device designed to monitor blood glucose levels in real-time and automatically regulate insulin for diabetes patients. Blood glucose prediction plays a crucial role in these systems by enabling proactive responses to glucose variations, thereby preventing risks such as hypoglycemia or hyperglycemia and assisting patients in managing their condition effectively. However, Continuous Glucose Monitoring (CGM) sensor data often contain significant sensor noise. Without effectively reducing the sensor noise, prediction accuracy can be severely compromised. Therefore, we first present a deep learning (DL) method for noise reduction in CGM data and, second, propose a long-term blood glucose prediction approach based on the system response function, utilizing a multi-input(e.g., blood glucose, carbohydrate (CHO) intake, and insulin). In this study, simglucose, based on the UVA-PADOVA simulator, was utilized to test and evaluate the proposed methods. As a result, we found that noise reduction using deep learning (DL) was significantly more effective than conventional filtering methods. Furthermore, the proposed long-term blood glucose prediction approach reliably tracked blood glucose fluctuations in custom scenarios and accurately predicted daily glucose patterns. Even in random scenarios, the proposed model accurately captured blood glucose trends, closely aligning with actual BG values and demonstrating remarkable performance.