Non-invasive blood glucose monitoring using PPG signals with various deep learning models and implementation using TinyML

Accurate and continuous blood glucose monitoring is essential for effective diabetes management, yet traditional finger pricking methods are often inconvenient and painful. To address this issue, photoplethysmography (PPG) presents a promising non-invasive alternative for estimating blood glucose levels. In this study, we propose an innovative 1-second signal segmentation method and evaluate the performance of three advanced deep learning models using a novel dataset to estimate blood glucose levels from PPG signals. We also extend our testing to additional datasets to assess the robustness of our models against unseen distributions, thereby providing a comprehensive evaluation of the models' generalizability and specificity and accuracy. Initially, we analyzed 10-second PPG segments; however, our newly developed 1-second signal segmentation technique proved to significantly enhance accuracy and computational efficiency. The selected model, after being optimized and deployed on an embedded device, achieved immediate blood glucose estimation with a processing time of just 6.4 seconds, demonstrating the method's practical applicability. The method demonstrated strong generalizability across different populations. Training data was collected during surgery and anesthesia, and the method also performed successfully in normal states using a separate test dataset. The results showed an average root mean squared error (RMSE) of 19.7 mg/dL, with 76.6% accuracy within the A zone and 23.4% accuracy within the B zone of the Clarke Error Grid Analysis (CEGA), indicating a 100% clinical acceptance. These findings demonstrate that blood glucose estimation using 1-second PPG signal segments not only outperforms the traditional 10-second segments, but also provides a more convenient and accurate alternative to conventional monitoring methods. The study's results highlight the potential of this approach for non-invasive, accurate, and convenient diabetes management, ultimately offering improved health management.