Adsorption kinetics feature extraction from breathprint obtained by graphene based sensors for diabetes diagnosis

Abstract

Abstract The correlation between blood glucose and breath acetone suggested by several studies has spurred the research community to develop an electronic (e-nose) for diabetes diagnosis. Herein, we have validated the in-house graphene based sensors with known acetone concentration. The sensor performances such as sensitivity, selectivity and stability (SSS) suggested their potential use in acquiring breath print. The 10% higher mean saturation voltage for 30 diabetic subjects ensured a discrimination accuracy of 65% with a positive correlation (r = 0.88) between biochemically measured and non-invasively estimated (glycated haemoglobin) HbA1c. For the improvement of classification rate, thirteen features associated with the adsorption kinetics were extracted from the breathprint from each of the three sensors. The features given as an input to the Naïve Bayes classification model fetched an accuracy of 68.33%. Elimination of redundant features by distinction index and one-R feature ranking algorithm results in Naïve Bayes algorithm with improved performances. The success rate has improved to 70% using the subset of features ranked by one-R algorithm. These results indicated the use of feature ranking algorithms and prediction models for the improvement in accuracy of our in-house fabricated graphene based sensors.