Determination of SWIR Features for Noninvasive Glucose Monitoring Using Machine Learning

The use of infrared (IR) light for noninvasive glucose monitoring is a potential solution to reduce infection-related mortality rate for diabetic patients. However, IR spans a wide band and the relevant wavelengths need to be chosen. This paper presents an automated and computationally efficient model, capable of examining a large number of wavelengths, to determine the suitable ones for monitoring, based on feature selection and other machine learning techniques. The study examined wavelengths from 1300-2600nm which cover the majority of short-wave infrared (SWIR) band. For reliable ground truth, two datasets, D1 and D2, were used with 100 observations and 1000 observations respectively. In term of optimal performance with limited time and computational resources, Sequential Forward Floating Selection (SFFS) technique was chosen as a core feature selection algorithm due to its high accuracy and reasonable speed. Classifiers based on Support Vector Machine (SVM) were used to evaluate the performance of selected wavelengths. Principal Component Analysis (PCA) was used to enhance the accuracy. Pipeline and nested cross-validation techniques were adopted to prevent information leakage and biased results. The proposed approach managed to reduce the number of wavelengths by 65% for D1 and 58% for D2 while achieving accuracy and f1 score above 90%, which are 10% higher compared to other work in the literature. The feature selection results also suggest that suitable wavelengths fall in the range 1600–2600 nm.