Machine learning models for non-invasive glucose measurement: towards diabetes management in smart healthcare

The patients of diabetes require to observe and control their glycemic profile through continuous glucose level monitoring. The blood glucose measurement is possible through invasive, minimally invasive and non-invasive methods. Invasive method is traditional method for instant glucose measurement where glucose is measured by taking blood samples from the body. However, the repeated finger pricking increases the risk of blood-related infections and trauma. Hence, the development of non-invasive real time device is essential for smart healthcare to manage glucose-insulin balance. The paper presents machine learning models for non-invasive glucose measurement. So, various machine learning algorithms including Logistic Regression, KNN, Gaussian Naive Bayes, Linear Regression, Multi-polynomial Regression, Neural Network, XGBoost, Decision Tree, Random Forest and Support Vector Machine are applied on two dataset which are PIDD (UCI repository) and iGLU dataset (iGLU device). The comparative analysis is carried out where accuracy, training time, recall, precision, f-1 score and AUC curve is measured for classification algorithms. For regression algorithms, measures like accuracy, Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are used for comparison purpose. Random forest with 84% accuracy and 68% recall, 76% precision and 72% f1-score for PIDD and Decision tree with 70% accuracy, 8% mean absolute error (MAE) and 8.5% root mean square error (RMSE) for iGLU dataset gives best results. Clark grid analysis has also been done where all the values fall under zone A which gives 100% accuracy and the device is useful for medication purpose. The proposed work has been also compared with similar methods and the proposed work has excellent results in terms of MAD, mARD, RMSE and AvgE. The device would be ideal as non-invasive solution for continuous glucose monitoring.