SCIENCE A N D TECHNIQUES

Olive Oil Quality Control by Using Pyrolysis Mass Spectrometry and

Artificial Neural Networks Royston Goodacre, Douglas B. Kell & Giorgio Bianchi

T he adulteration of food­stuffs is a major problem from several points of view which necessitat es

the implementat ion of rapid and accurate methods for its de tection. This art icle reports the application of Curie-point pyrolysis mass spec trom etry (PyMS) to the recognition of a variety of extra \irgin olive oil , and some of the same oils which had been deliberately adulterated with other lower­grade seed oils. Artificial neural networks (ANNs) were train ed successfully to dist inguish (the pyrolysis mass spectra of) extra virgin oils from adulterated olive oils. We consider that the combination of PyMS a nd ANNs co nst itutes a powerful and novel approach to the de­tection of adulterations in olive oil and food generally.

INTRODUCTION

There is an ongoing requ ire­ment for methods which might be exploited for the rapid char­acterisation of biological sys­tems, for instance in determin­ing whether a particular food­stuff has the provenance claimed for it or whether it has been adulterated . We here de­scribe a novel approach lo the solution of this problem, and its application to the distinction between extra virgin and adul­terated olive oils.

Pyrolys is is the thermal degradation of complex materi­al in an inert atmosphere or a vacuum. It causes molecules to cleave at their weakest points to produce smaller, volatile fragments called pyrolysate. A

36

mass spectrometer can then be used to separa te the compo­nents of the pyrolysa le on the basis of th e ir mass-lo-charge ratio (m/z) so as to produce a pyrolysis mass spectrum, which can then be used as a «chemical profile » or finger­print of th e complex material analysed, and the co mbined technique is then known as py­rolysis mass spectrometry (PyMS) (Meuzelaar et al. , 1982; Aries et al., 1986).

Virgin olive oil is the oil ex­tracted by purely mec hanical means from sound , r ipe fruits of the olive tree (Olea europaea L.). Such oils with a free fatty acid content (in terms of oleic acid) below I % are known as «extra virgin », whilst oils \Yith good flavour but grea ter addity are commonly called "virgin" or "ordinary". Lower grades in-

elude oils that have a high acid­ity and fl avou r and/or aroma defects which are corrected by refining; these are known as "lampante" oils. Oli ve oil con­tributes significantly lo the nu­tritiona l and health benefits of Mediterranean-type diets and, uniquely among vegetable oils, the fl avour of olive o il is best enjoyed without refining. Olive oil therefore commands a high­er price than do other vegetable oils, and these and other prop­erties mean that there is a grea t temptation lo adulterate olive oils with other seed o ils (Kirit­sakis and Marakakis, 1991 ).

A number of me thods have been proposed for the detection of olive oil adulterat ion, for examp le c hromat ograp hic (Kapoulas and Passaloglou-Em­manouliclou, 1981 ), near in­frared (Salo et al., I 991) and

qw,i,~~~~~o:.:.ll:U.U""""' ts pyrolysis mass spectrometer rt-IE SA rrJ<;H LIBRARY

SCIENCE fl' ~EAfNf. f. OLIY JE/No.47 - JUNE 1993 AND 'N~OR r."·,\ I !ON S( iNiCf

Supplied by the British Library 19 Feb 2020, 16:05 (GMT)

I . I

i

SCIENCE A N D TECHNIQUES

8

.... A r:: 6 = Q

CJ r:: .s -;: 4 Q .... ~ ~ ! 2 ~ ... ~

0 60 80 100 120 140 160 180 200

Mass

8

.... B' r:: = 6 Q u

= .s ~ 4 s "'"' Q

~ ! 2 = Q,)

~ ~

0 60 80 100 120 140 160 180 200

Mass

Figure 2. Normalised pyrolysis mass s11ectra of an extra virgi11 olive oil (A) ad11lremted with 5% s1111jlower oil ( B ).

UV spec troph o to metri c meth­ods (Passaloglou-Emmanouli­dou , 1990) , al though none a p­pears to have found widespread usage or all-round effec tiveness .

Chemometrics is the di sc i­pline concerned with the appli ­ca tion of s tatistical and mathe­matical methods to chemi ca l data (Massart et a l., 1988). A r elated approach is the use o f (artifi c ia l) neura l networks (ANNs), which can be consid ­ered as collecti ons of very sim­ple «computational units» whi ch can take a num eri ca l input (pyrolys is mass spec tra )

OLIY /E/No.47 - JUNE 1993

and transform it into an output (i.e., whether the oil is of virgin qua lity o r adultera ted) (Rumel­ha rt e t a l., 1986) . Th e great power of neural ne t works s tern s from the fac t th a t it is possib.le to «t rain » them. When tra ining is completed the ANN ma y then be exposed to «Un­known » pyrolysis m ass spectra and will th en «immediately» output th e s ta tus of the oils. If the outputs from the previously unkn own inputs a re accurate, the trained ANN is said to have genera lised.

~· ,- _ .. _. . . - - · ~· . :- ~· '";--· · . -~

Supplied by the British Library 19 Feb 2020, 16:05 (GMT)

METHODS AND RESULTS

Th e protocol of the experi ­m ent was as follows (Goodacre et a l., 1992 ). Two se ts of sam ­ples were prepared, one consist­ing o f 12 sa mples of variou s exlra virgin olive oils and the o th er of 12 samples va riously adulterat ed with 5-50% of soya , sunflower, pea nut , corn or re­fined olive oil. The experim ent was perfo rmed doubl e-blind , such that th e id entiti es of the second se t were not known. PyMS was perform ed at 530°C using a Hori zon Instrnments machine (see Fi g. 1) as de­scribed by Ari es e l a l. (1986) ; two typi cal sp ectra are shown in Fig. 2, wh ere it is obvious that it is not a t all easy to dis­tinguish them by eye. ·

ANN a na lyses were carri ed out using a user-fri endly, neural network s imula tion progra m , NeuralDesk (Fig. 3), which runs u nder Microsoft Windows/3.1 o n an IBM -compa tible PC . To ensure m aximum speed , we used a n accelerator board for the PC (N euSprint), based on the AT&T DSP32C chip and ob­ta in ed from the sam e source, whi ch effec ts a speed enhance­ment of some 100-fold, permit­ting the analysis (and updating) of so me 400,000 weights per second. We trained an ANN consis ting of an input layer of the 150 normalised averaged ion intensiti es from the pyroly­s is mass Sl)ec t ra of th e first se t of oils wit 1 mass range 51-200, and 1 hidden laye r of 8 nodes, using the standard back-propa­gation algorithm (Rumelhart et a l. , 1986) , coding virg in oils as 1, non-virgin as 0. The effective­ness of tra ining was expressed in term s of th e root mean squ a red (RMS) error between the ac tual and the desired out­puts over the entire tra ining set. When training had ceased (i.e., an RMS error of 0.001 bad been attained), the trained ANN was interroga ted with the nor­malised averaged ion intensities of th e pyrolys is mass spectra from the second set of oils (the unknowns). When the code was broken , it tran spired that the ANN had correc tly assessed each o il. In a typical run, the

37

SCIENCE A N D TECHNIQUES

NcuOc1k · 111111\N.OSK

fllc ltlifndow Uclp

Max Idle 1-::',------1 0

Conllol Window

Onandnmloc

0 '1 '' ... .

0 0 •

learning

OuidProp. Sl1ndard Beck Pro .

Figure 3. The Neural Computer Sciences Neura/Desk nrtificia l neural 11ef\Vork si11111latio11 progmm.

virgins were assessed wilh a code of 0 .99976 ± 0.0001 46 , (ra nge 0.99954 - 1.00016) and the non virgins with a code of 0.001079 ± 0.002838 (range 0.00026 ± 0.01009).

CONCLUSIONS AND ADVANTAGES OF THE

METHOD

PyMS has major advantages of speed (the typi ca l sample tim e is less than 2 mins) a nd a u tomati on, whi ch a llows ap­prox imately 300 samples to be analysed da ily. Furtherm ore, a fter the initi al outlay of some £50,000 on the British-m ade sys tem, running costs are rela­tive ly cheap, typica lly about £ I p er sample. The discriminating power of the approach is enor­mous, since if each (n or ­malised) mass is accura te to within 10% there are 101 50 possible pyrolysis mass spectra, and all biological material may be pyrolysed in the way de­scribed. The s tudy repor ted here clearly shows that the use of pyrolysis mass spectrometry and artifici al neura l networks was able rapidly and accurately to assess the contamination of virgin olive oils with 5-50% corn , peanut , soya, sunflower oil, or refined olive oil. In con­clusion, the combina tion of

38

th ese two techniques cons ti ­tutes a novel, powerful and uni ­versa l approach to the detection of food adul tera tion .

ACKNOWLEDGEMENTS

This work is supported by the Biotec hnology Direc tora te o f the UK SERC, under the terms of the LINK scheme in Bio­chemi cal E ngineerin g, in col­labora tion with Horizon Instru­ments, Neura l Computer Sci­ences and Zeneca p ie.

The Centra l Offi ce of In fo r­mation of the Crown is thanked fo r permission to rep rodu ce this article.

Royston Goodacre-Douglas B Kell Depar1111enr of Biological Sciences,

University of Wales, Abe1)'.mvyth, Dyfed S423 3 DA U. K.

Giorgio Bianchi Istir11to Sperimcntale per 111 Elaiotccnica

Co11trada "Fonte Umana", Pescara, /Joly

Supplied by the British Library 19 Feb 2020, 16:05 (GMT)

BIBLIOGRAPHY

Aries, R. E., Gutteridge, C. S. & Ott­ley, T. W. Evaluation of a low-cost, automated pyro lysi -mass spec­trometer, J . Anal. Appl. Pyrol. 9, 81-98 ( 1986).

Goodacre, R., Kell, D. B. & Bianchi , G. Neura l networks and oli ve oil. Nature 359, 594 ( 1992).

l<apoulas, V. M. & PassaJog lou-Em­manoulidou, S. Detection of adul ­tera ti on of oli ve oil with seed oil s by a combination of co lumn and gas liquid chromatograph y. J . Amer. Oil Chem. Soc. 58. 694-697 (198 1).

Kiritsakis, A. & Marakaki s, P. Oli ve oil anal ys is. In: Essential Oils and Waxes, eds. H.F. Linskens and J.F. Jackson 1-20. (S pringer, Heidel­berg, l 991 ).

Massa rt , D. L. , Yandeginste, 8 . G. M., Deming, S. N., Michotte, Y. & Kaufman, L. Chemometri cs : A tex tbook. (E lsevier, Amsterdam, 1988).

Meuzelaar, H. L. C., Haverkramp, J. & Hil eman, F. D. Pyrolysis mass spec trometry of recent and fossil biomateri als. (Elsevier, Amster­dam, 1982).

Passalog lou-Emmanoulidou. S. A comparati ve stud y of UV spec­trophotometric methods for detec­tion of oli ve oil adulteration by re­fined oil s. Lebensmittel Untersuch. Forsch. 19 1, 132- 134 ( 1990).

Rumelhart, D. E. , McClelland , J. L. & the PDP Research Group . Paral­lel Distributed Processing. Experi­ments in the Microstructure of Cognition. (MIT Press, Cambridge, 1986).

Sato, T., Kawano, S. & Iwamoto, M. Nea r infrared spec tral patterns of fa tty ac id analysis from fat s and oils. J. Amer. Oil Chem. Soc. 68. 827-833 (1991).

* * *

OLIY tE/No.4 7 - JUNE 1993

T r

! " ·I

l

1 1 I l l . I . f

· l I

·1 I

I . J

l I j

I f

'j J

) J J

'I 'l .l l