A Non-Invasive Blood Glucose Monitoring for Diabetics With Breath Biomarkers: An Ensemble-of-Classifiers Model

Abstract

As a supplemental technique for disease diagnosis, expelled breath monitoring is becoming extremely prevalent. Nevertheless, because of the significant number of factors that must be considered, researchers are forced to create new algorithms for accurate data interpretations. In our study, a novel noninvasive breath biomarkers-based diabetic detection strategy is introduced by upcoming three main processes “pre-processing, feature extraction and diabetes detection”. Raw data (breath signal) are initially pre-processed via data normalization to increase acquired breath signal quality. Most important features, such as “Improved empirical wavelet functions (IEWF), empirical mode decomposition (EMD), local mean decomposition (LMD), continuous wavelet transform (CWT) method, and entropy-based feature,” are extracted from the pre-processed data. These attributes are fused together and are subsequently used to train the ensemble of classifiers in the diabetes detection stage. An ensemble of classifiers, including an “optimized quantum deep neural network (QDNN), a convolutional neural network (CNN), and a recurrent neural network (RNN)”, is used to depict the diabetes detection phase. The CNN and RNN are trained using the retrieved fused features. The optimized QDNN receives the output of the CNN and RNN. The final detected result is provided by the optimized QDNN. A unique hybrid optimization model, Archimedes Principle with Enhanced Honey Badger Exploitation (APEHBE) algorithm, that combines the Archimedes Optimization Algorithm (ArchOA) and the Honey Badger Algorithm (HBA) is used to fine-tune the weight of QDNN. The effectiveness of the anticipated model is then verified by a comparison analysis.