Using Machine learning to predict blood glucose level based on Photoplethysmography

Photoplethysmography (PPG) is a non-invasive technique used to monitor the tiny changes in blood vessels caused by the heartbeat. This study obtained PPG signals from a commercial PPG module, followed by Signal Quality Index (SQI) detection to assess the signal quality and further signal processing, such as filtering and baseline shift. As a result, a total of 36 features were extracted from PPG signals in both time and frequency domains, and a correlation matrix was used to inspect the correlation among features. Machine learning models, including Random Forest (RF) and eXtreme Gradient Boosting (XGBoost), were employed to train and predict blood glucose levels. Of the 36 features, 9 were selected for use as input data for red and infrared light signals. Four datasets, namely “Red”, “Infrared”, “Composite” (a combination of “Red” and “Infrared”), and “Modified Composite” (a refined version of the “Composite” dataset that reduces feature collinearity and enhances prediction accuracy), were investigated. The RF model trained with the “Modified Composite” dataset yielded the best prediction, with a Mean Absolute Relative Difference (MARD) of 5.15% and an R-value of 0.93.