Estimating Blood Glucose Levels Using PPG Signals: A Deep Learning Approach with Diverse Patient Profiles

Accurate and continuous monitoring of blood glucose levels is crucial for effective diabetes management. Photoplethysmography (PPG) has emerged as a promising modality for non-invasive blood glucose monitoring. In this study, we investigate how blood glucose levels can be measured using PPG signals. To do this, we analyze a largescale dataset from VitalDB, which includes information from 6,388 subjects. By including diverse patient profiles in our dataset, we ensure that our models can be applied to a wide range of individuals. To capture the complex patterns and relationships in PPG signals, we employ a complicated deep learning approach based on convolutional neural networks (CNNs). Our models have achieved a mean absolute error (MAE) of 23.99, demonstrating their accuracy and reliability. Furthermore, we evaluate the clinical accuracy of the predictions using the Clarke Error Grid analysis, which reveals a high agreement within clinically acceptable ranges. By utilizing the comprehensive VitalDB dataset, our study contributes to a deeper understanding of PPG-based blood glucose prediction and validates the effectiveness of our models on a large-scale cohort. Additionally, we perform a comparison with recent papers, demonstrating why our model excels in various aspects such as model generality with a large dataset, high clinical accuracy, and the requirement of only 10 seconds of PPG data for effective prediction. The results highlight the potential of PPG signals as a valuable tool for non-invasive blood glucose monitoring.