Intelligent estimation of blood glucose level using wristband PPG signal and physiological parameters

• Design of a non-invasive blood glucose estimation system using photoplethysmography (PPG) signal and physiological parameters. • The Mel frequency cepstral coefficients (MFCC) of the PPG signal are used as features. • The extreme gradient boost regression (XGBR) model achieves a high correlation coefficient for the predicted glucose values. • The feature selection based on XGBR provides the best 5 features. • The model with selected 5 features improves the value of the correlation coefficient. • The XGBR model with the 5 features achieves clinically acceptable results with less complexity. The inconvenience and risk associated with the regular use of invasive blood glucose measurements has led to tremendous research in this area. This paper proposes the design of a non-invasive blood glucose estimation system using novel Mel frequency cepstral coefficients features of wristband photoplethysmogram signal and physiological parameters. A dataset from 217 participants of a hospital in Cuenca Ecuador is used to validate the proposed model. The support vector regression (SVR) and extreme gradient boost regression (XGBR) techniques are used to estimate blood glucose levels (BGL). The XGBR technique achieves the least value for the standard error of prediction (SEP), 9.78 mg/dL. Further, 5 features are selected from the feature set based on the feature importance in XGBR. The XGBR model with the reduced feature set results in further reduction of SEP value (5.53 mg/dL) with a correlation coefficient of 0.99. Standard Clarke error grid analysis and Bland-Altman analysis shows that the predicted glucose values are in the clinically acceptable region. The results of the proposed model demonstrate the potential of wearable BGL monitoring technology.