Independent component analysis-processed electronic nose data for predicting Salmonella typhimurium populations in contaminated beef

• Published in: Food control 2008-03, Vol.19 (3), p.236-246
• Main Authors: Balasubramanian, S., Panigrahi, S., Logue, C.M., Doetkott, C., Marchello, M., Sherwood, J.S.
• Format: Article
• Language: English
• Subjects: ambient temperature; beef; Blind source separation; Electronic nose; food contamination; food pathogens; Food safety; food storage; Independent component analysis; inoculum density; loins (meat cut); Meat contamination; microbial detection; plate count; predictive microbiology; principal component analysis; Salmonella typhimurium; storage time
• ISSN: 0956-7135; 1873-7129
• DOI: 10.1016/j.foodcont.2007.03.007

The changes in the headspace from stored beef strip loins inoculated with Salmonella typhimurium and stored at 20 °C were detected using an electronic nose system. Once the data was obtained six area-based features were extracted from the collected sensor data pertaining to the six metal oxide sensors present in the electronic nose. These extracted features were next dimensionally reduced by principal component analysis (PCA) and the independent components (IC) were extracted by FastICA package. The extracted independent components and principal components (PC) were compared by plotting them individually against the Salmonella population counts. A stepwise linear regression prediction model with the IC and PC as inputs was also built. The prediction model with IC as input performed better with an average prediction accuracy of 82.99%, and root mean squared error (RMSE) of 0.803. For the model using the PC as the input, the average prediction accuracy was 69.64% and the RMSE was 1.358. The results obtained suggest that the use of higher-order statistical techniques like ICA could help in extracting more useful information than PCA and could help in improving the performance of the sensor system. Further analysis needs to be carried out on larger datasets, and by using non-parametric data analysis techniques like artificial neural networks to build the prediction models from the ICA extracted components.