Augmented Noninvasive Glucose Monitoring Via Photoplethysmography and Particle Swarm Optimization-Based Random Forest Regression

Abstract The increasing prevalence of diabetes highlights the need for accurate, efficient, and noninvasive blood glucose monitoring. Conventional methods are often invasive, costly, and carry a risk of infection. To address these limitations, this study presents a noninvasive blood glucose detection system that integrates photoplethysmography (PPG) with advanced signal processing and machine learning techniques. Wavelet transform was employed for denoising, while median filtering corrected baseline drift, jointly enhancing the stability and quality of the PPG signal. A random forest regression (RFR) model optimized using particle swarm optimization (PSO) was developed to improve the accuracy and robustness of glucose level predictions. The wavelet-based denoising process effectively stabilized signal amplitude within the range of 0.3 V to 1.3 V. The PSO-optimized RFR model demonstrated strong predictive performance, achieving a coefficient of determination (CD) of 0.889, with corresponding mean absolute error (MAE) of 0.21 mmol/L, mean squared error (MSE) of 0.12 mmol/L, and root mean squared error (RMSE) of 0.33 mmol/L. These results confirm the model's reliability and precision. By combining PPG with a PSO-enhanced machine learning model, this study offers a promising approach to noninvasive glucose monitoring and lays a strong foundation for future health monitoring technologies and early clinical interventions.