Comparison of Noninvasive Blood Glucose Estimation Using Various Regression Models

Abstract

In this work we aim to evaluate the performance of three types of machine learning models implemented for blood glucose level (BGL) estimation. Pulse photoplethysmography signal is acquired from 611 human subjects and used for the analysis. Time and frequency domain features are extracted using (1) Frame based and (2) Single Pulse Analysis technique. These two features are used as input to train neural network, Support vector machines and Random forest models. These trained models are used for the estimation of BGL values. The BGL estimation performance of these models, (i) neural network, (ii) SVM, (iii) RF and (iv) K-fold RF are compared based on two feature sets i.e. Frame based time and frequency domain features and Single Pulse Analysis based time and frequency domain features. The performance of each system model is evaluated on the basis of, (i) Coefficient of determination i.e. R2, (ii) Spearman’s coefficient of correlation, (iii) Pearson’s coefficient of correlation and (iv) Clarke error grid analysis. We observed that Single Pulse Analysis technique shows better performance as compared to Frame based technique. The highest R2 value (0.95) for Single Pulse analysis is obtained for K-fold RF network. For Single Pulse Analysis, all other models also show comparable BGL estimation accuracy with R2 values ranging from 0.91 to 0.94. According to Clarke error grid analysis the values that lie in class A and class B are clinically accepted. We obtained highest prediction accuracy for Single Pulse analysis with K-fold random forest with 93.2% (class A) and 6.8% (class B).