Non invasive blood glucose estimation using green light photoplethysmography and machine learning

Non-communicable diseases, such as diabetes, are the leading cause of mortality worldwide. Effective diabetes management is crucial for ensuring the well-being of diabetics. Existing glucose monitoring technologies are often invasive and uncomfortable, eliciting anxiety among patients. Non-invasive procedures offer a promising solution for these issues, but their widespread adoption is restricted by cost and accuracy constraints. This study investigates the use of green light Photoplethysmography (PPG) signals for non-invasive blood glucose monitoring. A custom-designed PPG acquisition setup was developed to collect PPG data from 80 subjects under controlled conditions. Simultaneously, reference capillary blood glucose readings were obtained using a lancing device to serve as the gold standard. Signals were enhanced by applying different processing techniques and 32 features were extracted, which were scaled and subjected to correlation analysis to retain the highly correlated features. Feature engineering further optimized the feature set, which was then used to train and validate regression models. Model performance was evaluated using R 2 (coefficient of determination), mean absolute error (MAE), and bias analysis across glucose ranges. Among the tested models, the Random Forest Regression (RFR) showed the best performance with an R 2 value of 0.92 and MAE of 4.8 mg/dL. Predicted glucose levels demonstrated minimal bias across glucose ranges, with mean differences of −5.11 ± 0.80 mg/dL (