Giuseppina Amato visited Wallon Agricultural Research Centre

Purpose of the visit

The purpose of this visit was to test the chemometrics classification techniques in the procedure for detection and identification of the presence of mammalian Meat and Bone Meal (mMBM) in ruminant feeds even in the presence of fish meal in the same feed, in order to identify the different species. Furthermore, this visit allowed the fellow to get trained in the application of multivariate statistics to assess results obtained from a large number of variables, in order to find models that allow prediction of the animal species.

Description of the work carried out during the visit

In order to achieve these main objectives, the fellow participated in a current study at CRA-W that dealt with the development and validation of new methods for detection of low-level MBM in feedstuffs and development of the models to predict the identity of animal species.
In this study, near infrared microscopy (NIRM) was used to identify different classes of animal proteins. Samples of fish meals (n=10) and meals of land-animal origin (n=50) were ground. All samples was analysed using an Auto Image Microscope connected to Fourier transform near infrared spectrometer (FT-NIR; PerkinElme, Walthman, MA, USA). Spectra were obtained from the ratio between raw spectra and the background, consisting of the measurement of the Spectralon. PerkinElmer Autoimage software 3.2.1 was used for spectrum collection and storage.
Statistical analysis of the obtained results was performed using MATALAB v. 7.0 (The Mathwork Inc., Natick, MA, USA). The Support Vector Machines (SVM) algorithm was used to construct models to identify class origin. SVM analysis can perform binary classification; it learns to discriminate between the members and the non-members of a class. After learning the features of the class, the SVM recognises unknown samples as a member of specific class.
The constructed SVM models was validated in order to estimate the misclassification rate. Two different procedures were selected for model validation. Firstly, the set of available input-output measurements was divided in two parts, one part for training and one for testing, thus enabling comparison of the performance of different models constructed using the training set, on the test set. The second validation procedure used was the well-known leave-one-out cross-validation, that consists of splitting the learning data set, comprising L patterns, into a training set of size L-1 and a test of size 1 and averaging the squared error on the left-out pattern over the L possible
ways of obtaining such a partition. Both validation procedures were used in this study and final model was constructed using all data and reporting statistics for both procedures.

Description of the main results obtained

The total number of spectra for all samples was 7259. The Principal Component Analysis (PCA) was performed in order to reduce the database to an homogeneous data set. The final selected data set consisted of a calibration set with 1380 spectra, of which 420 were fish spectra, and a validation set comprising 630 spectra, of which 90 were fish spectra. The calibration set was used to construct the fish vs non-fish equation.
The model yielded a classification rate of 100% (based on calibration set) and had a prediction ability which reached an average correct classification of 95% and 95.5% for leave-one-out cross-validation ad for the test set respectively. In both cases, the results show that a clear discrimination is possible between fish meal and other species, given the high success rate obtained in both cases.Several independent sets were also predicted with external validation. The particles were detected as fish or as MMBM for all samples after application of two equations (fish vs non-fish and MMBM vs non-MMBM).

The results show that NIR microscopy combined with SVM can be used for detection of MMBM with high success rate. In contrast to optical microscopy, this method offers one main advantage: it is not dependent on the subjectivity of the analyst because particles identified form their NIR spectral fingerprint and not by visual inspection.
Furthermore, this STSM has allowed that the fellow got trained in the application of several multivariate statistic techniques, which were new for her. This will give her the possibility to apply this techniques in her home institution in future research. Future perspectives: During this STSM detection and identification of the presence of mammalian Meat and Bone Meal (mMBM) in feeds even in the presence of fish meal in the same feed, was obtained. In the field of fish meal and as well as the feed contamination, the case of sea mammals material need to be addressed in the near

Future collaboration with host institution

After the STSM, the collaboration with the host institution is expected to be continued later in 2010 after return of the fellow to Italy. A further occasion for implement this collaboration will be the CRL meeting that will be organised on April 28-29th 2010 in Turin at IZS.

Projected publications/articles resulting or to result from the STSM
It is planned to disseminate the present and future results at least in a conference abstract and/or paper. Possible publication in a peer-review journal will be also explored. The COST action Feed for Health FA0802 will be acknowledged in the resulting publications.

Giuseppina Amato STSM
Design & Programming: Zenlex