Machine Learning for Metal Oxide Gas Sensor Analysis

Semiconductor metal oxide sensors have picked up traction in the science community because of their ability to detect extremely low concentrations of volatile organic compounds (VOC) in air. Application of these sensors as a diagnostic tool for cancer are of interest because of their potential for early detection. Patients with cancer exhale certain VOC biomarkers in their breath. A low-cost, non-invasive method for early diagnosis may be through detection of these biomarkers via metal oxide sensors. Although sensors have already been developed that detect cancer, the type of cancer is still variable. A method of differentiating one cancer from another is by identifying the concentrations of the gases prevalent in a patient’s breath. In this project we investigate various methods for qualitative and quantitative methods for gas characterization. Datasets are taken from the University of California Irvine, and it is found that principal component analysis is a valid method for visualizing sensor effectiveness. Algorithms such as support vector regression and random forest are valid if predicting new values within the training set. Attempts to predict outside of data range result in inaccurate predictions. Mixed gas quantitative predictions were also attempted and were inaccurate in predicting due to overlap in sensor responses. Principal component analysis was however successful when visualizing the separation between pure gas and mixed gas concentrations.