<![CDATA[Conventional blood glucose measurement devices remain dominated by invasive methods that require direct contact with body tissues and cause patient discomfort, motivating the development of accurate and non-invasive blood glucose monitoring approaches. Therefore, blood glucose estimation has performed using Photoplethysmography (PPG) signals by exploiting the Velocity Photoplethysmogram (VPG) and the Acceleration Photoplethysmogram (APG) as derivative signals. PPG signals were acquired non-invasively from 58 participants using an optical MAX30102 sensor placed on the finger. Morphological fiducial features and Mel-Frequency Cepstral Coefficients (MFCCs) were extracted from PPG, VPG, and APG signals to capture temporal, morphological, and spectral characteristics related to peripheral hemodynamic changes. The extracted features were then used to model blood glucose levels using a Random Forest algorithm. The results showed that the combined use of PPG, VPG, and APG produced the best performance, achieving Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Relative Difference (MARD) values of 6.35 mg/dL, 9.74 mg/dL, and 6.07%, respectively. Clarke Error Grid Analysis (CEGA) showed that 97.44% of the predictions were located in Zone A and 2.56% in Zone B, indicating that all estimates were within clinically acceptable regions. The main scientific contribution of this study lies in integrating derivative PPG signals with fiducial and MFCC-based feature representations within a unified machine learning framework. These findings indicate that combining PPG, VPG, and APG provides a richer physiological representation than PPG alone and offers practical potential for the development of wearable-based, non-invasive blood glucose monitoring technology.]]>
