Variable selection towards classification of digital images: identification of altered glucose levels in serum

Abstract

classed as <100 mg/dL, and altered levels are classified as pre-diabetic (100–125 mg/dL) or diabetic (>125 mg/dL). Herein, we propose a method to identify control, pre-diabetic, or diabetic simulated and real-world samples based on their glucose levels using classification-based variable selection algorithms [successive projections algorithm (SPA) or genetic algorithm (GA)] coupled to linear discriminant analysis (SPA-LDA and GA-LDA) towards analyzing red–green–blue digital images. Images were recorded after glucose enzymatic reaction, whereby 250 μL of reactant content of samples were captured by using a common cell phone camera. Processing was applied to the images at a pixel level, where 72.2% of the pixels were correctly classified as control, 79.2% as pre-diabetic, and 90.9% as diabetic using SPA-LDA algorithm; and 76.8% as control, 81.4% as pre-diabetic, and 91.7% as diabetic using GA-LDA algorithm in the validation set containing nine simulated samples. Eight real-world samples were measured as an external test set, where the accuracy using GA-LDA was found to be 92%, with sensitivities ranging from 70% to 100 and specificities ranging from 90% to 99%. This method shows the potential of variable selection techniques coupled with digital image analysis towards blood glucose monitoring