Human breath-print identification by E-nose, using information-theoretic feature selection prior to classification

The composition of bodily fluids reflects many aspects of health status of a patient. Breath is another sample that may be useful for diagnosis of infectious and other diseases. Analysis of breath has the advantage of being less invasive than analysis of other fluids such as blood and bronchial biopsy. Two recent studies, using either mass spectrometry or electronic nose (E-nose) technologies, showed there are definite “breath-prints” that characterised individuals despite temporal variation in internal metabolism and environment. In this study we demonstrate that by employing an information-theoretic feature selection method that is specific to the problem together with machine learning techniques, we can dramatically improve (cross-validated) identification of individuals through their breath using a very small selected subset of E-nose measurement features. Indeed, we demonstrate here that we can identify the 10 individuals in this study with perfect accuracy using fewer than 10 features.