Non-invasive Blood Glucose Detection System with Infrared Pulse Sensor and Hybrid Feature Neural Network

The rising prevalence of diabetes increases the demand for daily blood glucose detection, necessitating the urgent development of non-invasive blood glucose detection systems. To enhance the convenience of blood glucose monitoring, we employed Infrared Pulsed Sensing (IPS) to capture photoplethysmography (PPG) signals. PPG signals effectively reflect changes in blood volume within the human body, providing rich information about blood glucose. In this paper, we designed a blood glucose detection system based on an IPS and a deep hybrid feature neural network. By deploying IPS and deep learning algorithms on a Raspberry Pi, the system is equipped with data collection, analysis, prediction, and display capabilities for real-time blood glucose monitoring. Previous studies relying on shallow machine learning for blood glucose detection struggled to capture the complex underlying correlations between PPG signals and blood glucose. Our proposed deep hybrid feature neural network model, DCC-Net, achieves end-to-end blood glucose prediction by capturing different temporal and spatial features in PPG signals. The F2M fusion module combines multi-scale and multi-level features, effectively modelling the intricate non-linear relationship between PPG signals and blood glucose, thereby improving classification performance. We recruited 290 participants and used IPS to collect PPG signals, constructing an IPS-PPG dataset. Testing DCC-Net on the IPS-PPG dataset yielded an impressive overall classification accuracy of 0.92, demonstrating its high accuracy. Furthermore, our model showed good generalisation performance on the publicly available dataset PPG-BP. The proposed approach exhibits significant potential for advancing research and development in non-invasive BG detection systems.