Blood glucose prediction by breath analysis system with feature selection and model fusion

It has been shown that the concentration of acetone in breath is correlated with the subject's blood glucose level (BGL). Therefore, noninvasive BGL monitoring of diabetics can be achieved by the analysis of components in breath. In this paper, a breath analysis device with 10 gas sensors is designed to measure breath samples. Transient features are extracted from the signals of the sensors. Sequential forward selection is applied on the features to find the most informative ones. In order to reduce the interference brought by the inter-subject variance of breath acetone, global and local BGL prediction models are built and fused. The two models are based on different training strategies and have different advantages. Experiments were conducted using 203 breath samples from 36 diabetic subjects. Results show that the accuracy of the proposed feature is better than other similar features and the model fusion strategy is effective. The mean absolute error and mean relative absolute error of the system are 2.07 mmol/L and 20.69%, respectively.