PREDICTING ORGANOLEPTIC SCORES OF SUB-PPM FLAVOR NOTES - PART 2 - COMPUTATIONAL ANALYSIS AND RESULTS

In Part 1 of this paper (T. C. Pearce and J. W. Gardner, Analyst, 1998 , 123, 2047 we describe a novel method for predicting the organoleptic scores of complex odours using an array of non-specific chemosensors. The application of this method to characterising beer flavour is demo nstrated here by way of predicting a single organoleptic score as defi ned under the joint EBC/ASBC/MBAA international flavour wheel for beer . An experimental study was designed to test the accuracy of the odour mapping technique for this prediction of organoleptic scores of added reference compounds within a chemically complex lager beer background . Using the flow injection analyser (FIA) system comprising 24 conduct ing polymer sensors, also described in Part 1, sampling was conducted on spiked lager beers. A dimethyl sulfide spike was added at the 20-80 ppb v/v level to simulate a range of organoleptic scores (0-5.5 out o f 10) for flavour note no. 0730-''cooked vegetable''. A certain amount of sensor drift was observed over the 12 d testing period which is sh own to account for significant variance in the data-set as a whole. Th e effect of the sensor drift was reduced by applying a linear drift mo del, which may be generally applied when the between-class variance du e to the difference in odours is small when compared with the within-c lass variance due to the drift, which increases approximately linearly over time. Careful use of this drift compensation model, coupled with judicious selection of pre-processing and pattern recognition techniq ues, maximised the between-class variance and so improved the overall classification performance of the system. After applying detailed expl oratory data analysis, statistical, and neural classifier techniques, the organoleptic score was predicted with an accuracy of +/-1.4 tout o f 10) and 95% confidence. Our results show that it is possible to gene rate subjectively defined organoleptic flavour information, using mult i-sensor arrays and associated data-processing that is comparable in a ccuracy to sensory and GC-based techniques.