Exhaled breath signal analysis for diabetes detection: an optimized deep learning approach.

In this study, a flexible deep learning system for breath analysis is created using an optimal hybrid deep learning model. To improve the quality of the gathered breath signals, the raw data are first pre-processed. Then, the most relevant features like Improved IMFCC, BFCC (bark frequency), DWT, peak detection, QT intervals, and PR intervals are extracted. Then, using these features the hybrid classifiers built into the diabetic's detection phase is trained. The diabetic detection phase is modeled with an optimized DBN and BI-GRU model. To enhance the detection accuracy of the proposed model, the weight function of DBN is fine-tuned with the newly projected Sine Customized by Marine Predators (SCMP) model that is modeled by conceptually blending the standard MPA and SCA models, respectively. The final outcome from optimized DBN and Bi-GRU is combined to acquire the ultimate detected outcome. Further, to validate the efficiency of the projected model, a comparative evaluation has been undergone. Accordingly, the accuracy of the proposed model is above 98%. The accuracy of the proposed model is 54.6%, 56.9%, 56.95, 44.55, 57%, 56.95, 18.2%, and 56.9% improved over the traditional models like CNN + LSTM, CNN + LSTM, CNN, LSTM, RNN, SVM, RF, and DBN, at 60th learning percentage.