Lightweight Time–Frequency Neural Network for Non-Invasive Glucose Estimation on Wearable Devices

Abstract

Continuous glucose monitoring is essential for effective diabetes management, but most clinically validated systems rely on minimally invasive sensors that require subcutaneous insertion, periodic calibration, and frequent replacement. These limitations create a critical need for non-invasive glucose estimation methods that are accurate, wearable, and deployable on resource-constrained devices. In this study, we propose a lightweight neural network architecture for non-invasive glucose estimation using biomedical signals, including PPG, ECG, and accelerometer (ACC) signals. To address the non-stationary nature of physiological signals, we employ discrete cosine transform (DCT) representations that preserve both temporal and spectral characteristics. The model uses a multi-branch convolutional architecture with anisotropic kernels that independently capture temporal dynamics and spectral correlations, enabling efficient feature extraction while having only 20.5 k parameters. Evaluation on two public datasets demonstrates robust cross-modal performance, achieving over 98 % of predictions within zones A+B of the Clarke Error Grid. These results show that compact neural networks can enable accurate, real-time, non-invasive glucose monitoring on wearable devices.