Precise Detection and Quantitative Prediction of Blood Glucose Level With an Electronic Nose System

Abstract

Blood glucose level is an important health indicator. Non-invasive, easy-to-use glucose detection and monitoring methods and tools are desperately needed, especially for patients with diabetes. In this work, we developed a new method to quantitively identify and analyze the blood glucose level by measuring the biomarkers in breath with an electronic nose (E-Nose) system based on a metal oxide (MOX) gas sensor array. Advanced machine-learning models have been studied and developed to precisely predict the blood glucose level based on the measurement of 41 participants for 10 days. The testing result shows that the E-Nose system and proposed analysis models identify blood glucose levels at an accuracy of 90.4% and a small average error of 0.69 mmol/L in blood glucose concentration. This study indicates that the E-Nose system enabled with machine learning is an efficient and precise method to achieve low-cost and non-invasive disease diagnosis.

Noninvasive Diabetes Detection through Human Breath Using TinyML-Powered E-Nose

“…However, precision alone does not account for false negatives, indicating individuals with diabetes that the model fails to identify. Thus, while high precision ensures correct positive predictions, it may not be sufficient if the model has a high rate of false negatives, potentially leading to the omission of people with diabetes mellitus [ 44 , 45 , 46 , 47 , 48 , 49 ].…”

“…Firstly, the sensitivity of the implemented MQ gas sensor series to high RH levels poses a significant challenge, necessitating continuous validation and calibration of the medical devices to detect errors in the system’s operation promptly. This is crucial, especially considering the limited lifespan and susceptibility of MOS sensors to %RH [ 16 , 24 , 35 , 45 , 46 , 48 , 49 ]. Additionally, the need for a dehumidifier to clean the e-nose sample chamber adds complexity to sensor maintenance and operational logistics.…”

“…Subsequently, feature scaling was conducted using Z-score normalization for CO, alcohol, acetone, ketones, humidity, and temperature sensor readings; this process ensured all signals were standardized onto a similar scale, exhibiting properties akin to a normal distribution, thereby facilitating comparison and analysis within the study. Equation (7) demonstrates Z-score normalization, also known as standard normalization [ 35 , 44 , 49 ]. where represents the original sample value, is the sample mean, and is the sample standard deviation.…”

“…Previous clinical studies in diabetes mellitus employed algorithms based on VOC features selection and extraction using Principal Component Analysis (PCA) [ 38 , 44 , 45 , 46 , 47 , 48 ]. Regression models in breath analysis provide quantitative BGL results with 90.4% accuracy [ 49 ].…”

Prediction of Individual Dynamic Thermal Sensation in Subway Commute Using Smart Face Mask

“…The laboratory testing in our previous experiment [ 40 ] found that EBT is significantly correlated with TSVs. Besides predicting individual thermal sensations in a subway, the researchers were trying to leverage the use of face masks by integrating sensors to monitor SKT, EBT, respiratory patterns, CO 2 monitoring, biomarkers for inflammation, and airborne pathogen detection [ 57 , 58 , 59 , 60 , 61 ].…”