A Stratified Modeling-Machine Learning Approach to Improve the Accuracy of Non-Invasive Blood Glucose Estimation Using Photoplethysmography Signals

Abstract

Diabetes is a silent killer that can only be controlled with continuous monitoring of blood glucose levels. The method commonly used is invasive and has various weaknesses, but it is more accurate than non-invasive methods. This research aims to develop a method to increase the accuracy of non-invasive estimation of blood glucose levels using photoplethysmography (PPG) signals. The proposed method is to carry out stratified modeling-machine learning. The tested classifiers were support vector machines (SVM), KNN, Naïve Bayes, decision tree, and neural network. The prediction model used simple linear, logarithmic, second-order polynomial, exponential, and power regression. Applying stratified modeling using linear regression in the non-diabetes stratum and logarithmic regression in the diabetes stratum obtained a mean absolute relative difference (MARD) value of 4.5%, root mean square error (RMSE) of 18.9 mg/dl, Pearson correlation 0.985 and Clarke error grid analysis (CEGA) 96% in region A and 4% in region B. The implementation of stratification reveals a marked improvement in efficacy, manifested as a reduction in the MARD by 77.83%, a decrease in the RMSE by 51.91%, an enhancement in the Pearson correlation by 0.065, and a CEGA by 100% in regions A and B, thereby being clinically acceptable. Implementing a stratified modeling-machine learning approach can improve the accuracy of non-invasive blood glucose level estimates.