Photoplethysmography based non-invasive blood glucose estimation using systolic-diastolic framing MFCC features and machine learning regression.

Introduction: Accurate and non-invasive blood glucose estimation is essential for effective health monitoring. Traditional methods are invasive and inconvenient, often leading to poor patient compliance. This study introduces a novel approach that leverages systolic-diastolic framing Mel-frequency cepstral coefficients (SDFMFCC) to enhance the accuracy and reliability of blood glucose estimation using photoplethysmography (PPG) signals. Methods: The proposed method employs SDFMFCC for feature extraction, incorporating systolic and diastolic frames. The systolic and diastolic points are identified using the Savitzky-Golay filter, followed by local extrema detection. Blood glucose levels are estimated using support vector regression (SVR). The evaluation is performed on a dataset comprising 67 raw PPG signal samples, along with labeled demographic and biometric data collected from 23 volunteers (aged 20 to 60 years) under informed consent and ethical guidelines. Results: The SDFMFCC-based approach demonstrates high accuracy (99.8%) and precision (0.996), with a competitive root mean square error (RMSE) of 26.01 mg/dL. The Clarke Error Grid analysis indicates that 99.273% of predictions fall within Zone A, suggesting clinically insignificant differences between estimated and actual glucose levels. Conclusion: The study validates the hypothesis that incorporating a new framing method in MFCC feature extraction significantly enhances the accuracy and reliability of non-invasive blood glucose estimation. The results highlight that the SDFMFCC method effectively captures critical physiological variations in PPG signals, offering a promising alternative to traditional invasive methods.