An LSTM-Based Model for Non-invasive Blood Glucose Prediction Utilizing BVP Signals

Diabetes is a severe and long-lasting metabolic condition, and the effective management depends on continuous monitoring and control of blood glucose levels. However, traditional invasive methods, such as fingertip blood sampling, cause inconvenience and discomfort to patients, resulting in reduced compliance and motivation for diabetes management. Furthermore, these methods increase the risk of infection due to frequent blood sampling. Therefore, this research aims to develop a non-invasive model for predicting glucose levels based on blood volume pulse (BVP) signals. These devices employ Photoplethysmography (PPG), a technique that monitors changes in blood volume due to periodic heartbeats, to acquire BVP signals. By leveraging machine learning, specifically the long short-term memory network (LSTM), our model thoroughly processes, cleans, segments, and extracts features from these signals. Our approach enables not only real-time monitoring of blood glucose levels but also the prediction of future glucose trends, crucial for timely medication adjustments and disease management. Experimental results reveal that our LSTM model significantly outperforms traditional RNN and linear regression models in predicting blood glucose levels within clinically acceptable ranges. This study offers a more convenient and accurate method for diabetic patients to monitor their blood glucose levels, thereby enhancing their overall health management.