Ultrasound-driven deep learning for glucose monitoring in flowing blood

Abstract

Diabetes management requires frequent blood glucose monitoring, yet current methods remain invasive and inconvenient. We present a novel non-invasive approach for classifying blood glucose levels using ultrasound and deep learning. The proposed method employs a single-element ultrasound transducer to capture acoustic signals from flowing whole blood, which are then analyzed by a convolutional neural network (CNN) to determine the glucose concentration category. This approach combines ultrasound's non-invasive blood glucose monitoring capabilities with CNN pattern recognition to achieve high classification accuracy without preprocessing blood samples. In contrast to prior techniques, our approach can analyze unprocessed whole blood in real time. We validated the system on blood samples spanning a wide range of glucose concentrations. Experimental results demonstrate that the CNN can reliably distinguish multiple clinically relevant glycemic ranges directly from the raw ultrasound waveforms. The key advantages of this method are its non-invasive nature, the high accuracy enabled by artificial intelligence (AI)-based signal analysis, and the capability to operate on whole blood directly. This integrated ultrasound & CNN-based glucose classification system promises a convenient, needle-free solution for diabetes monitoring.