Pii: S0956-5663(99)00035-4

Two series of experiments are reported which result in the discrimination between Helicobacter pylori and other bacterial gastroesophageal isolates using a newly developed odour generating system, an electronic nose and a hybrid intelligent odour recognition system. In the first series of experiments, after 5 h of growth (37°C), 53 volatile ‘sniffs’ were collected over the headspace of complex broth cultures of the following clinical isolates: Staphylococcus aureus, Klebsiella sp., H. pylori, Enterococcus faecalis (10 ml), Mixed infection (Proteus mirabilis, Escherichia coli, and E. faecalis 3×10 ml each) and sterile cultures. Fifty-six normalised variables were extracted from 14 conductive polymer sensor responses and analysed by a 3-layer back propagation neural network (NN). The NN prediction rate achieved was 98% and the test data (37.7% of all data) was recognised correctly. Successful clustering of bacterial classes was also achieved by discriminant analysis (DA) of a normalised subset of sensor data. Cross-validation identified correctly seven ‘unknown’ samples. In the second series of experiments after 150 min of microaerobic growth at 37°C, 24 volatile samples were collected over the headspace of H. pylori cultures in enriched (HPP) and normal (HP) media and 11 samples over sterile (N) cultures. Forty-eight sensor parameters were extracted from 12 sensor responses and analysed by a 3-layer NN previously optimised by a genetic algorithm (GA). GA-NN analysis achieved a 94% prediction rate of’ ‘unknown’ data. Additionally the ‘genetically’ selected 16 input neurones were used to perform DA-cross validation that showed a clear clustering of three groups and reclassified correctly nine ‘sniffs’. It is concluded that the most important factors that govern the performance of an intelligent bacterial odour detection system are: (a) an odour generation mechanism, (b) a rapid odour delivery system similar to the mammalian olfactory system, (c) a gas sensor array of high reproducibility and (d) a hybrid intelligent model (expert system) which will enable the parallel use of GA-NNs and multivariate techniques. © 1999 Elsevier Science S.A. All rights reserved.