Abstract Diabetes mellitus is a chronic disorder characterized by persistent hyperglycemia that damages multiple organs. With global prevalence rising, accurate, convenient, and non-invasive glucose monitoring is urgently needed. However, current diagnostic methods—such as point-sample tests and continuous glucose monitoring (CGM)—remain limited by invasiveness and potential inaccuracy. In this study, we conduct an in vitro feasibility evaluation toward non-invasive glucose monitoring using an 80 MHz high-frequency ultrasound (HFU) transducer paired with a convolutional neural network (CNN). Signals were converted into time-frequency representations and analyzed using CNNs to classify blood samples by glucose concentration. Despite the inherent heterogeneity and noise in whole blood, the system achieved approximately 68% multi-class accuracy across glucose levels. These results indicate that ultrasound-based, AI-driven signal analysis offers a promising alternative for continuous, non-invasive glucose assessment, supporting improved diabetes management in clinical and home settings.