Estimation of blood glucose by non-invasive method using photoplethysmography

This paper presents a system which estimates blood glucose level (BGL) by non-invasive method using Photoplethysmography (PPG). Previous studies have shown better estimation of blood glucose level using an optical sensor. An optical sensor based data acquisition system is built and the PPG signal of the subjects is recorded. The main contribution of this paper is exploring various features of a PPG signal using Single Pulse Analysis technique for effective estimation of BGL values. A PPG data of 611 individuals is recorded over duration of 3 minutes each. BGL value estimation is performed using two types of feature sets, (i) Time and frequency domain features and (ii) Single Pulse Analysis (SPA). Neural network is trained using above mentioned proposed feature sets and BGL value estimation is performed. First we validate our methodology using the same features used by Monte Moreno in his earlier work. The experimentation is performed on our own dataset. We obtained comparable results of BGL value estimation as compared with Monte Moreno, with maximum R2 = 0.81. Further, BGL estimation using (i) Time and frequency domain features and (ii) Single Pulse Analysis (SPA) is performed and the resulting coefficient of determination (i.e., R2) obtained for reference vs. prediction are 0.84 and 0.91, respectively. Clarke Error Grid analysis for BGL estimation is clinically accepted, so we performed similar analysis. Using Time and frequency domain feature set, the distributions of data samples is obtained as 80.6% in class A and 17.4% in class B. 1% samples in zone C and Zone D. For Single Pulse Analysis technique (SPA) the distribution of data samples are 83% in class A and 17% in class B. The proposed features in SPA have shown significant improvement in R2 and Clarke Error grid analysis. SPA technique with the proposed feature set is a good choice for the implementation of system for measurement of non-invasive glucometer.