Extraction and selection of parameters for evaluation of breath alcohol measurement with an electronic nose

The ethanol concentration in realistic breath samples was analyzed using an electronic nose. Conditions were selected so that the samples would reflect those collected in a real drunk driver situation. Hence, parameters such as intake of food and beverage, tobacco habits, as well as the order of participating volunteers were allowed to be variable. The setup was unexpectedly robust towards inter- and intrapersonal variations in breath samples as well as long-term variations. The standard error (16 mol ppm) was the limiting factor but the statistical detection limit was well below 0.1 ppm. The standard error corresponds to between 9% (Austria) and 36% (Sweden) of lowest legally accepted levels. Even though this is regarded as a significant error, there are several options of optimization. Incorporating feature extraction and forward selection together with artificial neural network for prediction of the ethanol concentration showed, besides increasing the accuracy, to be a valuable tool generating feedback of possible improvements of the sensor array.