Noninvasive Blood Glucose Estimation Using Pulse Based Cepstral Coefficients

In this work we aim to investigate the importance of cepstral coefficients (CC) of Photoplethysmograph (PPG) signal in the estimation of noninvasive blood glucose levels (BGL). Cepstral features are widely used in speech signal processing applications such as robust speech recognition and speech synthesis. We recorded PPG signal of diabetic and non-diabetic subjects. We computed 1) frame based and 2) single pulse based cepstral coefficients of PPG signal to estimate BGL values. The performance of the frame based and single pulse based technique using CC features for BGL estimation are compared based on four performance metrics namely 1) Coefficient of determination i.e. R2, 2) Spearman's and 3)Pearson coefficient of correlation and 4) Clarke error grid analysis. We found that Cepstral features based on single pulse technique outperforms frame based technique in terms of above mentioned performance metrics. We obtained highest R2, Spearman and Pearson coefficient values of 0.90, 0.94, and 0.95 respectively. We also implemented Clarke error grid analysis which is clinically accepted method in BGL estimation. Using Single Pulse technique we obtained 85.2% BGL values in Class A and 13.6% values in class B, where estimation in both classes are clinically accepted. in class B, where estimation in both classes are clinically accepted.