DELIVERABLE - phasmafood.eu · Project Full Title: Portable photonic miniaturised smart system for on-the-spot food quality sensing ... (iii) detection of food fraud. In this report,

This document is part of a project that has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 732541. It is the property of the PhasmaFOOD consortium and shall not be distributed or reproduced without the formal approval of the PhasmaFOOD Management Committee.

Project Acronym: PhasmaFOOD

Grant Agreement number: 732541 (H2020-ICT-2016-1 - RIA)

Project Full Title: Portable photonic miniaturised smart system for on-the-spot food

quality sensing

DELIVERABLE

Deliverable Number D3.2 Deliverable Name Feasibility results and use case

benchmarking-v2 Dissemination level Public

Type of Document Report

Contractual date of delivery 30th June 2018

Deliverable Leader AUA

Status & version V4.0

WP / Task responsible WP3 / T3.1, 3.2 and 3.3 - AUA

Keywords: alcoholic beverages, fish, fraud, fruits, grains, image analysis,

meat, milk, mycotoxins, nuts, oil, spectrum analysis, spoilage,

vegetables

Abstract (few lines): The PhasmaFOOD project aims to develop a miniaturized, multi-

parameter and programmable sensing node for (i) detection of

mycotoxins, (ii) detection of food spoilage and prediction of

shelf-life; and (iii) detection of food fraud. In this report, an

updated feasibility assessment of the above mentioned (sub-)

use cases is displayed. D3.2 is the second report of results on

real-life samples in a series of successive reports (D3.1 - M9,

D3.2 - M18 and D3.7 - M27) leading to a final feasibility

assessment of the proposed use cases. In order to do so, this

deliverable describes the standard operating procedures for the

PhasmaFOOD sensors, sampling strategies, data assessment and

conclusions until M18 of the project. For the mycotoxin use case

Project Title: PhasmaFOOD Contract No. 732541 Project Coordinator: INTRASOFT International S.A.

Page | 2

D3.2 – Feasibility results and use case benchmarking-v2, Version: v4 – Final version, Date 28/06/2018

(in particular aflatoxin B1), first results on the sub-use cases of

maize flour and almonds are reported. For the spoilage use

case, extensive experimental data have been produced which, in

addition to fish (presented, in part, in D3.1), refer to minced

pork, rocket, baby spinach and pineapple. Finally, for the third

use case on food fraud, initial feasibility assessment of the usage

of the PhasmaFOOD sensors in detecting skimmed milk powders

(SMP) is reported, and experiments on minced meat

adulteration are described.

Deliverable Leader: George-John E. Nychas (AUA)

Contributors: AUA, CNR, DLO, IPMS, UTOV

Reviewers: IPMS, WINGS

Approved by: INTRA


Page | 3


Document History

Version Date Contributor(s) Description

0.1 05/04/2018 AUA, CNR, DLO,

IPMS

Early version-Request for partners’

contribution

1.0 27/05/2018 AUA, CNR, DLO,

IPMS, UTOV

First draft version for review/update by

contributors

2.0 06/06/2018 AUA, CNR, DLO,

IPMS, UTOV

Second draft version for consortium

review

3.0 18/06/2018 AUA, CNR, DLO,

IPMS, UTOV

Third (final) draft version for internal

review

4.0 28/06/2018 AUA, CNR, DLO,

IPMS, UTOV Final version


Page | 4


Table of Contents Definitions, Acronyms and Abbreviations 8

1 Executive summary 10

2 Introduction 11

2.1 State of the art 11

2.1.1 Use case 1: detection of mycotoxins in food products 11

2.1.2 Use case 2: detection of early sign of spoilage, spoilage and shelf-life estimation in

meat, fish, fruit and vegetables. 13

2.1.3 Use case 3: food fraud 14

2.2 Description of micro-sensors 15

2.2.1 NIR microspectrometer 15

2.2.2 VIS spectrometer 17

2.2.3 CMOS camera 17

2.3 WP3 and D3.2 scope, strategy and planning 18

3 Materials and methods 20

3.1 SOPs for micro-sensor operation 20

3.1.1 NIR spectrometer 20

3.1.2 VIS spectrometer 23

3.1.3 CMOS camera 25

3.2 Detection of mycotoxins 25

3.2.1 Maize flour 33

3.2.1.1 Materials 33

3.2.1.2 Experimental design 35

3.2.1.3 Implementation of sensors (food science laboratory) 35

3.2.1.4 Other experimental procedures 35

3.2.2 Milk powder 36

3.2.3 Paprika powder 36

3.2.4 Tree nuts 36

3.2.4.1 Materials 36




Page | 5



3.3 Detection of food spoilage and shelf-life prediction 39

3.3.1 Meat 39




3.3.2 Fish 42




3.3.3 Fruit and vegetables 44




3.3.3.4 Pilot measurements 45

3.4 Detection of food fraud 46





3.4.2 Meat 48




3.4.3 Alcoholic beverages 49




3.4.4 Edible oils 50

4 Data analysis 51

4.1 Mycotoxin detection 51


Page | 6



4.1.2 Tree nuts 53

4.2 Food spoilage detection and shelf-life prediction 53

4.2.1 Meat 53

4.2.2 Fish 53


4.3 Food fraud detection 54


4.3.2 Meat 53



5 Results and discussion 56

5.1 Feasibility of micro-sensor application for detection of mycotoxins 56



5.1.3 Paprika powder 59

5.1.4 Tree nuts 59

5.2 Feasibility of micro-sensor application for detection of spoilage 65

5.2.1 Meat 65

5.2.1.1 Microbiological data 65

5.2.1.2 Multispectral imaging (MSI) sensor (VideometerLab system) 66

5.2.1.3 FTIR spectroscopy 67

5.2.2 Fish 68


5.2.3.1 Microbiological data 69

5.2.3.2 Spectral data 70

5.3 Feasibility of micro-sensor application for detection of fraud 73

5.3.1 General food test 73


5.3.3 Meat 75



Page | 7



6 Conclusions and outlook 76

References 77


Page | 8


Definitions, Acronyms and Abbreviations Acronym Title

CFU Colony Forming Units

CO Confidential, only for members of the consortium (including Commission

Services)

CR Change Request

D Demonstrator

DL Deliverable Leader

DM Dissemination Manager

DMS Document Management System

DoA Description of Action

Dx Deliverable (where x defines the deliverable identification number e.g. D1.1.1)

EIM Exploitation Innovation Manager

EU European Union

FLUO Fluorescence

FM Financial Manager

FTIR Fourier-transform Infrared

MAP Modified Atmosphere Packaging

MSI Multispectral Imaging

MSx project Milestone (where x defines a project milestone e.g. MS3)

Mx Month (where x defines a project month e.g. M10)

NIR Near-infrared

O Other

P Prototype

PC Project Coordinator

PLS Partial Least Squares

PM partner Project Manager

PO Project Officer

PP Restricted to other program participants (including the Commission Services)

PU Public

QA Quality Assurance

QAP Quality Assurance Plan

QFD Quality Function Deployment

QM Quality Manager

R Report

RE Restricted to a group specified by the consortium (including Commission Services)

RMSE Root Mean Square Error

STM Scientific and Technical Manager

TL Task Leader

TVC Total Viable Counts

UV Ultraviolet


Page | 9


VIS Visible

WP Work Package

WPL Work Package Leader

WPS Work Package Structure


Page | 10


1 Executive summary The PhasmaFOOD project aims to develop a multifunctional optics-based miniaturized sensor

for fast characterization of foods by industry and consumers. The PhasmaFOOD scanner will be

smart-phone operated and consist of a visible (VIS) and near-infrared (NIR) spectrometer and a

board-level camera, together with relative illumination sources. Specifically, WP3 is focusing on

assessing the performance of the individual micro-devices (sensors) on three use cases, namely

(i) mycotoxin detection, (ii) food spoilage detection and shelf-life prediction, and (iii) food fraud

detection. This assessment will be performed prior to the delivery of the fully integrated

prototype device by (a) benchmarking individual food sub-use cases, (b) developing smart signal

processing in tandem with chemometrics, (c) building a basic reference database for each sub-

use case backed up by chemical and/or other fingerprinting reference methods, (d) validating

the established reference database, and finally (e) developing smart data correlation

algorithms between the tested micro-devices prior to full integration.

In this second version of reports dealing with the deliverable ‘Feasibility results and use case

benchmarking’, enriched content and updated (compared to D3.1) results of the feasibility of

the individual PhasmaFOOD sensors on the different target foods as well as benchmarking of

the individual food sub-use cases are reported. Specifically, in this deliverable, measurement

and sampling strategies, reference methods and chemometric protocols are elaborated upon

during the first 18 months of activity in the project.

The newly provided in D3.2 content expands to all three studied use cases. More specifically,

for the first use case, the new content comprises measurements of mycotoxins in grained maize

and almonds case using visible reflectance, fluorescence and NIR spectroscopy and image

acquisition and analysis. In the use case of spoilage and shelf-life prediction, extensive

additional data and/or data analysis results are reported on fish, minced pork, fruit (pineapple)

and vegetables (rocket, baby spinach) using visible reflectance, fluorescence and/or NIR

spectroscopy measurements, as well as relatively more well established spectral techniques

(vibrational spectroscopy and surface chemistry) as a comparison. Finally, for the third use case

on food fraud, initial feasibility of the usage of the FLUO, VIS and NIR spectra in skimmed milk

powders (SMP) is reported. This comprises an approach which authenticates SMP independent

of the fraud issue (non-hazardous low-value fillers and food safety related issues like chemical

nitrogen enhancers) using a one-class modelling approach. Experiments conducted under the

third use case and assessing minced meat adulteration also are described.


Page | 11


2 Introduction The need for the development of analytical techniques and/or instruments capable of providing

credible estimates of food safety and quality in a timely, rapid and non-invasive manner has

been well acknowledged during the last decade. Such a need has provided the trigger for truly

ambitious research attempts aiming at the development of sensors that will be able to

accurately describe various characteristics or properties of food products that delineate their

safety and/or quality status. Spectroscopy and imaging technologies hold a prominent position

among the approaches that have been evaluated and utilized for the purpose of food sensor

development, demonstrating a promising potential with regard to the assessment of various

aspects pertinent to food protection.

Therefore, the aim of PhasmaFOOD project is to develop a multifunctional optical sensing node

for food applications that will be ultimately operated by consumers. The scanner array consists

of a visible (VIS) and near-infrared (NIR) spectrometer and a board level camera (CMOS),

together with relative illumination sources, and will be smartphone operated. In this report, an

updated (compared to the D3.1 report) feasibility assessment of the three use cases on which

the PhasmaFOOD sensor will be tested is provided: (I) mycotoxin detection, (II) food spoilage

detection and shelf-life prediction and (III) food fraud detection. An extensive description of

each use case is contained in Deliverable Report D1.1 – Use case description and validation

plan. Each use case consists of different target foods and, therefore, different experimental

approaches and criteria for feasibility are used for each use case. This report (D3.2) is the

second report of results on real-life samples in a series of successive reports (D3.1 - M9, D3.2 -

M18 and D3.7 - M27) leading to a final feasibility assessment of the proposed use cases. Some

of the information provided in this report and specifically, information pertinent to the

description of the individual micro-sensors and the standard operating procedures for their

utilization in the conducted experiments, has been also provided in D3.1; nonetheless, the

inclusion of this information in D3.2 has been done for document coherency reasons.

2.1 State of the art

2.1.1 Use case 1: detection of mycotoxins in food products

Mycotoxins are toxic secondary metabolites produced by certain species of fungi mainly in

contaminated grains [1]. The presence of mycotoxins in animal feed and the food supply chain

constitutes a significant global food safety issue [2]. Indeed, consumption of food products

contaminated with mycotoxins has been associated with adverse health effects, ranging from

transient symptoms such as nausea and vomiting to long-term genotoxicity and carcinogenicity

[3]. One of the challenges in the detection of mycotoxins in food lies in the high cost, time and


Page | 12


labor requirements of well-established analytical methods such as thin layer chromatography

and liquid chromatography mass spectrometry [4].

Various alternative approaches have been recently assessed for their potential utilization in

mycotoxins’ detection, including the use of biosensors, electrochemical-based sensing

platforms, Fourier transform infrared (FTIR) spectroscopy, hyperspectral imaging and

complementary metal–oxide semiconductor (CMOS) sensor [5-11]. Nonetheless, given that

most of the analytical spectroscopic techniques that have been individually evaluated only

allow for the detection of rather high concentrations of mycotoxins, the conjunctional approach

of the PhasmaFOOD project is anticipated to result in more robust data, potentially allowing for

a higher accessibility of detection methods throughout the food supply chain. Established

mycotoxin contamination limits in different foods in the EU are summarized in Table 1.

Table 1: Mycotoxin contamination limits in different foods in the EU

(http://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:02006R1881-20140701).

Foodstuff

Max B1

alone

(μg/kg)

Max M1, B1,

B2, G1, G2

(μg/kg)

Almonds, pistachios and apricot kernels, intended for direct

human consumption or use as an ingredient in foodstuffs 8.0 10.0

All cereals and all products derived from cereals, including

processed cereal products, with the exception of foodstuffs

listed below

2.0 4.0

Maize and rice to be subjected to sorting or other physical

treatment before human consumption or use as an ingredient in

foodstuffs

5.0 10.0

Capsicum spp. (dried fruits thereof, whole or ground, including

chilies, chili powder, cayenne and paprika) 5.0 10.0

Given the established mycotoxin contamination limits, the measurement sessions in the current

part of the project are oriented to prepare samples useful to build calibration models

approaching legal limits on one hand, and collecting naturally contaminated samples to test

regression/classification models on the other.

In this context, the application of the PhasmaFOOD sensor is expected to constitute an

alternative means of mycotoxins’ detection, in a time-efficient, reagent-free, usable by non-


Page | 13


expert personnel and non-destructive to food samples manner. The different sub-use cases

considered in the PhasmaFOOD project are:

1. Maize (grained/flour): AF B1, total AFs and DON

2. Milk powder: AF M1 and M2

3. Paprika powder: AF B1 and total AFs

4. Tree nuts (homogenized, grained or whole nuts): AF B1 and total AFs

2.1.2 Use case 2: detection of early sign of spoilage, spoilage and shelf-life

estimation in meat, fish, fruit and vegetables.

Food quality, a central theme in food science research, has been considered as a term not easily

definable scientifically and that it comprises many different aspects, with the latter being

subject to constant changes. It has been opined that “food quality represents the sum of all

properties and assessable attributes of a food item”, and that this is accomplished through

three categories of quality: sensory value, suitability value and health value [12]. In the context

of a holistic assessment of food quality, several additional (to the aforementioned) categories

of quality have been also taken into account, including notional, cultural, political and ecological

values of food [12]. Indeed, it is widely accepted that the consumers’ perception regarding food

quality is a very important parameter when assessing food spoilage and shelf-life. Food

spoilage, a complex ecological phenomenon which is underlain mainly by the biochemical

activity of microorganisms, is related mainly to the sensoric and suitability values of food

quality. The food quality changes composing spoilage are related to the metabolic activity of

certain groups of microorganisms, referred to as “specific spoilage organisms”, and the type

and availability of the required energy substrates in foods [13, 14]. Although numerous

methods (organoleptic, microbiological or physico-chemical) have been developed for the

purpose of food quality assessment [15, 16], the majority of them are time-consuming, labor-

intensive, destructive, and provide retrospective information. Hence, various novel analytical

approaches have been recently evaluated and proposed for the non-destructive and rapid

assessment of food spoilage.

Examples of such promising approaches include enzymatic reactor systems, sensor arrays (e.g.

electronic noses), spectroscopy methods (e.g. vibrational, NMR or mass spectroscopy

techniques), as well as imaging technology approaches [17-20]. By means of combining visible

reflectance, fluorescence and NIR spectroscopy (and potentially CMOS images), the

PhasmaFOOD sensor is expected to be effective in estimating spoilage and shelf-life of a fresh

product. Still, since food spoilage is a rather complex ecological phenomenon, it should be kept

in mind that spoilage prediction can be a fairly difficult task, also in the PhasmaFOOD project.

One has to take into account biochemical activity of specific groups of microorganisms (strongly

associated with the shelf-life of various food products), the evolution of these specific groups of

microorganisms evaluated in conjunction with the physical and sensory changes of the food


Page | 14


product (e.g., color, appearance, odor) etc. The different sub-use cases considered in the

PhasmaFOOD project are “fresh” products which are highly perishable and may pose a high risk

towards consumers for foodborne infection or intoxication:

1. Meat

2. Fish

3. Fruit and vegetables

With regard to the microbial levels being associated with a spoilage status delineating the end

of a food product’s shelf-life, these depend on the product itself as well as on the applied

storage conditions (i.e. temperature and atmosphere inside the packaging). In order for the

effect of the environmental parameters to be also taken into account, the experiments

conducted under this use case, and for the different sub-use cases being studied, were

designed so as to include product storage under different temperatures (both isothermal and

dynamic) and, wherever applicable, under different packaging conditions (i.e. aerobic and

modified atmosphere packaging). Overall, based on existent hygienic regulations and/or

research data available in the scientific literature, total mesophilic microbial populations

exceeding 107 CFU/g (or 7 log CFU/g) have been associated with unacceptable spoilage, and

thus shelf-life termination, for all the aforementioned food categories (i.e. sub-use cases) [18,

21, 22].

2.1.3 Use case 3: food fraud

Food fraud is a collective term referring to the “deliberate substitution, addition, tampering or

misrepresentation of food, food ingredients or packaging, or false or misleading statements

made about a product for economic gain”, as defined by the United States Pharmacopeia

Convention [23]. In Europe, consumers are protected by EC Regulation No. 178/2002,

underpinning the concept of informed consumer choice in the purchase of food. Unfortunately,

the number of food adulteration and fraud cases being unraveled in several EU member states

is rising. Since the conventional laboratory analysis methods for detecting fraud/adulteration

are laborious and expensive, the need for a smart non-invasive, rapid and, ideally, hand-held

device is eminent. In this framework, various analytical technologies have been recently

assessed for their efficacy in detecting food fraud/adulteration and, thus, for their potential

value in food authentication. Such analytical technologies include visible/near-infrared

spectroscopy, UV-VIS spectroscopy, use of compact digital camera as well as image analysis

approaches, while examples of food authentication applications being evaluated include edible

oil composition monitoring, detection of minced meat adulteration and fresh/frozen-thawed

fish discrimination [24-28].

Overall, the smart sensor-based system which is planned to be developed within the

PhasmaFOOD project, is anticipated to allow for the accurate assessment of all the

aforementioned food protection aspects through the utilization of relevant spectral and/or


Page | 15


imaging data, and as such to be of great value for practical application throughout the food

supply chain (food manufacturers, retailers, food service, consumers). The different sub-use

cases being considered are:

1. Milk powder

2. Meat

3. Alcoholic beverages

4. Edible oils

2.2 Description of micro-sensors

2.2.1 NIR microspectrometer

The PhasmaFOOD smart sensing system integrates a miniaturised NIR spectrometer to cover

the wavelength range from 1000 to 1900 nm (Fig. 1). The device was developed by partner

IPMS, with a size of the sensor head of 16 × 17 × 12 mm. The central active component inside

this device is a miniature optical grating, which oscillates resonantly at a frequency of ~ 100 Hz,

driven by electrostatic forces from a comb-like structure. This micro-electromechanical (MEMS)

component was fabricated in the IPMS clean room [29]. The assembly of the spectrometer also

took place at IPMS. IPMS provides this spectrometer as a demonstrator kit complete with read-

out electronics but without housing, as this is part of the integration work of WPs 2 and 5. Once

the PhasmaFOOD device is finished in WP6/ month 18, the miniaturized NIR spectrometer will

be available for testing in WP3. Further technical details are elaborated in deliverable report

D6.1.

Figure 1: MEMS, miniaturized version of the NIR spectrometer developed at IPMS.

In the meantime, for laboratory testing in the early stages of WP3, a more robust solution was

found, namely the NIR spectrometer SGS1900 (Fig. 2). This device incorporates the same

technology and measuring principle as the above miniaturized version and is commercially

available from Hiperscan GmbH, a spin-off of IPMS. For the purposes of WP3, in order to test

the general applicability of the sensing method for the three PhasmaFOOD use-cases, an

SGS1900 instrument was provided by IPMS as a free loan to PhasmaFOOD partners.


Page | 16


Figure 2: The NIR spectrometer SGS1900 used for testing of the general applicability of the sensing method for the PhasmaFOOD use cases.

The size of the SGS1900 housing is 105 x 80 x 86 mm. It comprises a MEMS-based scanning

grating for spectral dispersion, and an uncooled InGaAs diode for detection. A detailed data

sheet is available online (HiperScan). A halogen light source provided illumination (in-house

constructed by IPMS) was coupled via SMA 905 connector into an Ocean optics Y-shaped fibre

bundle (QR400-ANGLE-VIS). The fibre bundle comprises a probe tip with a window, which is set

at 30° to the front face of the fibres. During measurement, this window is placed in direct

contact with the sample. Light from the halogen source is fed to the sample and the diffuse

reflectance from the sample is collected back into a 400 µm core optical fibre in the centre of

the fibre bundle. This fibre then transmits the collected light to the SGS1900 spectrometer via

SMA connector (Fig. 3). The software “Quickstep” (version 0.99 by Hiperscan GmbH [30]) serves

as operation software.

Figure 3: Experimental set up of the NIR spectrometer compartments.


Page | 17


2.2.2 VIS spectrometer

The UV-VIS spectrometer used in the PhasmaFOOD sensor is the Hamamatsu C12880MA. The

device has a spectral range from 340 to 850 nm (larger than other models), high sensitivity

(Conversione Efficiency 50 μs/e-), compact dimensions (20 mm × 15 mm × 10 mm) and spectral

resolution of 15 nm. In our specific application, it is employed to detect the fluorescence signal

of aflatoxins and for visible range spectroscopy in the two other use cases, so only the visible

range is exploited. A UV filter with 400 nm cutoff wavelength is introduced in front of the

spectrometer aperture to avoid that the UV illumination used to excite the fluorescence could

saturate the detected signal. Such a filter will also be integrated into the PhasmaFOOD sensing

prototype, into the parallel beam section between sample and VIS spectrometer. In the first

phase of the project, the evaluation board from Hamamatsu has been used to drive the VIS

sensor. A custom developed board is under development by the project partners CNR and

WINGS (see D6.1).

Figure 4: Hamamatsu C12880MA UV-VIS spectrometer.

2.2.3 CMOS camera

A miniature CMOS camera from Ximea (MU9PC-MH model) has been chosen to be integrated in

the PhasmaFOOD sensor (Fig. 5). The CMOS camera measures 15 mm × 15 × 8 mm and has a

resolution of 5 mega pixels (2592 × 1944 pixels), which to the best of our knowledge fits the

sensing requirements of the use cases (D1.1) and hardware requirements (D1.2). The Ximea

camera has high definition, high sensitivity, low crosstalk and low noise image capture

capability in an ultra-small and lightweight design in the visible range. The CMOS camera may

help the user to identify signs of spoilage in food samples (use case 2), but also to identify

specific sample features that will drive the selection of the analysis model. For instance, one

such feature is granularity, since the dimension of particles in the case of grained cereals and

nuts may affect the spectral results (use case 1) and, hence, the choice of the analysis method

depends on this feature. In addition, the camera may be employed to perform multispectral

imaging by using Red-Green-Blue (RGB) imaging analysis and possibly LEDs of different colors in

order to get complementary information on the targeted use-case.


Page | 18


A second assembly, including VIS spectrometer and CMOS camera with relative temporary

illumination sources (3 x white 400 K light LEDs, 2 x 365-nm UV LEDs and a triple LED including

636-red, 565-green and 585-yellow) has been assembled by UTOV in cooperation with CNR and

used in the measurement sessions with maize and almond under use case 1.

Figure 5: The Ximea MU9PC-MH camera with a 20 Eurocent coin for size comparison.

2.3 WP3 and D3.2 scope, strategy and planning

The goal of this WP is to test the individual use cases as described in D1.1 for feasibility. This

includes first activities on practical application of the PhasmaFOOD micro-sensors on realistic

samples, exploration of the methods for smart signal processing, chemometrics, development

and adaptation of algorithms for decision making and initial validation of the spectral databases

belonging to each individual (sub-) use case. Furthermore, data-fusion strategies are explored

to fully exploit the advantage of using multiple spectral sensors and to accurately estimate the

feasibility of the use cases as they were proposed in D1.1. This deliverable provides an updated

and enriched (compared to D3.1) report on the feasibility of the use cases: detection of

mycotoxins (aflatoxin) in grains (maize) and tree nuts (almonds) (T3.1); detection of early sign

of spoilage, spoilage and shelf-life estimation in fruit, vegetables, meat and fish (T3.2); and

detection of food fraud in skimmed milk powder and meat (T3.3).

The results presented in this report were obtained by using the individual UV-VIS, NIR and the

CMOS camera sensors in the period M1-M18, with an emphasis being placed on the newly

obtained experimental and/or data analysis results (i.e. the ones generated in the period M9-

M18). This strategy has been adapted to already gain insight in the practical application of the

sensors before completion of the first PhasmaFOOD prototype. Furthermore, the findings in

this report provide feed-back on the design of the integrated PhasmaFOOD instrument and the

timely delivery of spectral databases which are extensive enough to provide TLR4 validated

results before the end of the project.

Within this reporting period (till M18), activities have been concerned with writing and revising,

when needed, standard operating procedures (SOPs) for standardized operation of the micro-


Page | 19


sensors in each laboratory where feasibility testing is performed (CNR, AUA and DLO).

Furthermore, for the period M9-M18, an additional planning has been produced for the

circulation of the micro-sensors between the involved laboratories, taking into consideration

the time required to perform measurements for feasibility estimation. Sensor time in the first

six months of Year 2 has been mainly allocated for studies under T3.2 (ca. M13-M15), due to

the lengthy duration of spoilage experiments, whilst for T3.1 (M15) and T3.3 (M15 and further

in Year 2) most samples can be acquired and stored until sensors are available. Therefore, in

this deliverable, first results on mycotoxins (maize flour and almonds, both naturally and

artificially contaminated) are reported, while for T3.2 additional (to those described in D3.1)

extensive data on spoilage of meat (minced pork), fruit (pineapple) and vegetable (rocket, baby

spinach) products are available. With regard to T3.3, initial feasibility of the usage of the

PhasmaFOOD sensors on skimmed milk powders is reported, while the experimental

procedures embraced for the purpose of assessing meat adulteration also are described.


Page | 20


3 Materials and methods

3.1 SOPs for micro-sensor operation

3.1.1 NIR spectrometer

In this section, the operation procedure for the Hiperscan SGS1900 NIR spectrometer (Serial

Number W1141 157) is described. The NIR micro-spectrometer will follow a similar operation

procedure albeit details like the software structure may differ (see D6.1). The steps for

operating the SGS1900 spectrometer include:

1. Prior to operation: Install the NIR acquisition software on your lab computer. In order to

run it, you require the MONO version 2.6.4 (or older). IPMS provides access to both

programs through a content server during the period of the PhasmaFOOD project.

2. Setting up: Connect the SGS1900 spectrometer to grid power (220 V) via the included

power adapter. The power status light of the SGS1900 should light up green. Connect

the SGS1900 spectrometer to your laboratory computer via the included micro-USB-to-

USB cable. Switch on your lab computer. Start the NIR acquisition software. The USB

status light of the SGS1900 should pulse in green during data transfers while the

instrument is being initialized. Once the instrument is properly initialized, you will see its

registration number in the “Device” window.

Inspect the two ends of the fiber bundle – one end comprises six fibers in a ring, the

other end comprises a single central fiber. Connect the single- fiber end to the SGS1900

spectrometer by plugging it into the SMA connector and fixing the screw. Connect the

six-fiber end to the halogen light source in the same way. Connect the light source to

grid power and switch it on by pressing the black switch at the back of the housing.

Looking into the angled window of the fiber probe, you should now see a ring of six

fibers lighting up. Wait for 5 minutes for the lamp to warm up.

3. Set number of averages: In the NIR acquisition software (“Properties” window), select

the number of averages. If this window does not show upon starting the software, you

can open it by choosing the “View” tab in the menu bar, then “Window layout” and then

selecting “Property editor”. A high number of averages results in longer measurement

times and in reduced data noise. A typical number is 1000 averages. Above 2000

averages, no further reduction of noise is expected. You may change the number of

averages by typing a number into the “Averages” window and pressing “Apply”. The

software will send a reminder message to renew the dark and background references

once the number of averages has changed.

4. Dark intensity reference (after each change of acquisition parameters): Make sure that

the SGS1900 spectrometer does not receive light by disconnecting the fiber from the


Page | 21


SGS1900 or by pointing the fiber probe towards dark empty space. In the software,

press the red button next to “Dark intensity”. Once the measurement is done, confirm

the setting by pressing “Apply”. The dark intensity measurement is required to correct

for the baseline noise of each measurement. Data can later be exported in a format that

has this baseline correction automatically included.

5. White background reference (after each change of acquisition parameters): Connect the

fiber bundle to SGS1900 spectrometer and halogen lamp as described above. Make sure

that the fiber probe is clean and dry. Leave the light on, hold the fiber probe so that it

touches a surface suitable as a white reference (a PTFE target or any other white

reference target suitable for NIR). In the software, press the red button next to

“Background”. The spectrometer should now record the spectrum of the halogen lamp

which is scattered back from the reference target. This measurement is saved internally

in the NIR acquisition software. Once the measurement is done, confirm the setting by

pressing “Apply”. The white background reference measurement is required to correct

for the lamp spectrum and calculate extinction/absorbance from the raw data.

Note: Dark intensity and background lamp spectrum may change slightly over time.

Therefore, each time the spectrometer is switched on, these references must be

recorded again. After that, press “Apply” and you’re ready to start measuring. If you are

measuring in long sessions, it is advisable to record new references approx. every 2

hours.

6. Measurement – setup: Measurements with the y-shaped fiber bundle are done in direct

contact with the sample. The tip of the fiber probe is cut at an angle of 30° in order to

avoid direct reflexes from the sample. Only diffuse scattered light is collected, which

contains information about the NIR absorption spectrum of the sample. Hold the fiber

probe so that the window at the tip of the probe and the sample surface are in full

contact, i.e. parallel.

Note: This measurement geometry is suitable for samples that absorb or scatter light in the NIR

spectral range. Transparent samples cannot be investigated this way. Intermediate

cases may arise when samples are physically thin such that the NIR lamp light is partly

transmitted through them. In these cases, please fold the sample over or stack several

samples to increase the thickness. A sufficient thickness is reached when no NIR lamp

light is observed through the back of the sample. If this precaution is not followed, the

physical background of the sample will influence the measurement.

7. Measurement – data acquisition: In the software, press the red button to acquire a

spectrum. Hold the fiber tip still during acquisition. Several spectra may be acquired in

sequence, e.g. to observe the variation across a single sample. For a quick overview of

the measured data, the acquisition software opens a tab with the spectrum diagram. Via

the “View” tab in the main menu, the “Plot type” view mode in this window can be

changed between intensity/absorbance/SNV and other modes. Under “Plot type” you

can select your preferred option. Default is an intensity plot corrected for the dark

intensity. For evaluation, Absorbance is preferred.


Page | 22


8. Saving data:

Option 1: Select the “File” tab in the main menu, then select “Export”. The acquisition

software will export all data from the currently viewed window in the currently used

View option into a single *.csv file. Caution: Please select Intensity mode before

exporting data. The acquisition software has a known bug in the export of absorbance

values. The dark counts offset is automatically subtracted in the saved intensity values.

The white reference spectrum required to calculate absorbance is not saved

automatically but is only stored in the software while running the spectrometer. In

order to calculate absorbance later during data analysis, the white lamp spectrum must

be recorded like a regular spectrum in the same diagram window. Tip: Measurements

can be grouped into sets by opening new diagram tabs, e.g. a new tab for each sample

or session.

Option 2: Select the “Extensions” tab in the main menu, then, “Acquisition”, then select

“Save after acquisition”. You will be asked for the file name. Please use a file name in

this format: “<sampleinfo>_01”. The software will then save all spectra in the currently

viewed diagram window in subsequently numbered (02, 03, etc. …) individual *.csv files.

These files contain a header with meta information like instrument settings and dark

counts, one column for x-axis (wavelength in nm), one column for intensity values (raw,

dark counts NOT subtracted), a third column for absorbance values (dark counts

subtracted and white reference spectrum taken into account). No bugs are known for

this absorbance calculation, i.e. the absorbance spectra can be trusted and directly used

in analysis.

It is recommended to use Option 2 for saving data as all meta-information and raw data

are automatically kept.

9. Shutting down: Close the acquisition software. Switch off the lamp. Disconnect all

electrical and optical connections of the lamp and SGS1900 spectrometer. Clean the

fiber probe window and put the plastic caps on the fiber ends/connectors of the

spectrometer, lamp and probe.

10. Cleaning: If necessary, in order to avoid contamination between measurements or

samples, the front window of the fiber probe can be cleaned by wiping with a simple

tissue containing ethanol, isopropanol or DI water. Please do not use any other solvent

(especially no acetone) or anything abrasive for cleaning. Please also do not dip the fiber

probe into any solvent for a longer time to avoid solvent entering the housing. Dipping

into liquids for a short time (minutes) is allowed, as long as the rim of the probe stays

above the liquid. The probe window should be cleaned and dried directly after the end

of the measurements, before the instrument is put away for the night. Simple optical

inspection by eye can tell whether the probe window needs cleaning between

measurements.


Page | 23


3.1.2 VIS spectrometer

The list of the possible steps for the operating procedure of the VIS spectrometer is described

below. Please note again that the spectrometer is protected with a UV excitation filter during

measurements. Therefore, strictly only VIS spectra will be obtained throughout the

PhasmaFOOD project. This does not affect the measuring procedure as it simply leads to zero

signal over the blocking range of the UV filter. Measured spectral data are considered valid in

the visible spectral range of the Hamamatsu spectrometer, i.e. between 400 nm and 850 nm.

Please note, that this restriction does not apply to the light sources, which will illuminate in

both, UV and VIS spectral ranges as the UV light is required to excite VIS fluorescence from the

sample. The following measuring procedure is based on the first integrated design of UV-VIS

spectrometer and lighting system (UV and white LEDs).

1. Power on the device (lights OFF).

2. Launch acquisition software.

3. From acquisition software, tool -> set parameters (integration time= for acquisition in

white light reflectance mode, according to use case, usually hundreds of microseconds,

number of average counts usually 10; for acquisition in fluorescence mode, integration

time in the range of hundreds of milliseconds and 3 average counts).

4. Dark calibration

a. While the lights are off, orientate the instrument toward a dark surface (the

optical spot) and keeping the device under a darkening hood start dark mode to

perform dark measurement and automatic storage.

5. White light reflectance mode (at the beginning of a new session or every day)

a. turn VIS light ON

b. Wait 15-20 minutes (this step depends on the white LED characteristics)

c. Point the acquisition window toward a white reference

d. Start Reference mode and perform acquisition of the white reference.

e. Put the acquisition window in front of the sample.

6. Measurement mode (performed for each single sample measurement)

a. Measurement mode and perform acquisition on the sample (different areas, and

different samples). Of course, acquisition parameters must be kept constant

throughout the measurement session.


Page | 24


b. The software for each measurement will give 3 files (*.csv format) as output:

sample1.dark, sample1.reference, sample1 (if not changed during the

measurement session, dark and reference file will be the same for all samples).

c. Data normalization and classification will be performed using the developed data

analysis models (that could be different for each case study) considering also the

dark and reference measurements.

7. 7. At the end of measurement session, perform measurement of the reference (using

measurement mode).

8. Close measurement software

a. Switch off illumination

b. Switch off device

For the Fluorescence mode

1-to 3 as above

4. Dark calibration

a. Direct the instrument toward a dark surface (the optical spot) and keeping

the device under a darkening hood and Start dark mode (UV lights off) to

perform dark measurement and automatic storage.

5. Reference acquisition

a. Turn UV light ON

b. Point the acquisition window toward a non-fluorescent reflective reference

standard

c. Start Reference mode and perform acquisition.

6. Start Measurement mode and perform acquisition on the sample (different areas, 5 spots

for each sample and different samples). Of course, acquisition parameters must be kept

constant throughout the measurement session. As already indicated, the software for each

measurement will give 3 files (*.csv format) as output: sample1.dark, sample1.reference,

sample1 (if not changed during the measurement session, dark and reference file will be the

same for all samples).

7. At the end of measurement session, perform measurement of the reference (using

measurement mode).

8. Close measurement software

a. Switch off illumination


Page | 25


b. Switch off device

The acquisition of both VIS and FLUO spectra from the same samples must be repeated in

different days during the same measurement session.

3.1.3 CMOS camera

Complete operating procedures for calibration and image analysis will be defined according to

use cases and upon completion of feasibility tests. In this phase of the project, the camera

inserted in the assembly together with the VIS spectrometer has been used in two modalities:

- white light modality (to extract sample morphological features like granularity) mainly

for use case 1 applications

- multi-color imaging using an array of monochromatic LEDs (UV, red, green, yellow) to

test the potentialities of multi-color imaging, possibly for use case 2 applications.

3.2 Detection of mycotoxins

In order to verify the feasibility of mycotoxins detection by fluorescence and VIS reflectance

spectroscopy, several tests have been performed on samples prepared by ISPA-CNR staff at

different aflatoxin B1 contamination levels. The samples have been prepared by the inoculation

technique described below and the corresponding contamination levels were in the range of

parts per million (ppm or μg/g) for the first sets of measurements in two different laboratories

(ISPA labs and Rikilt lab in Wageningen), and part per billion (ppb or ng/g) in the second set of

samples measured once (in ISPA labs), whereas another session has been repeated in M18 at

Rikilt lab in Wageningen. In this way, the ability of the VIS spectrometer to detect the presence

of the contaminants could be assessed. In the first set of samples, maize flour and almond flour

both with different granularity were analyzed. A reference sample for maize flour from Trilogy

Lab was also used and differences between non-contaminated and contaminated sample

spectra were recorded. Furthermore, samples from naturally contaminated maize were

characterized by standard laboratory techniques (mainly HPLC), and then VIS and FLUO spectra

were acquired and images were taken. In the second set of measurements, almond samples at

different contamination levels (0-291 ppb) were prepared and characterized. More accurate

calibration of the detectors will be performed in the successive project phases. Information

regarding the sample measurement sessions is provided in Table 2, while the mycotoxin

contamination levels in various samples are presented in Tables 3-10.


Page | 26


Table 2: Information regarding the sample measurement sessions.

Laboratory

location Samples

Contamination

range

Contamination

type Techniques Time

ISPA (Bari, IT) Maize,

Almond μg/g, ng/g (Trilogy)

Artificially, naturally (Trilogy)

VIS, FLUO M7

Rikilt (Wageningen, NL)

Maize, Almond

μg/g, ng/g (Trilogy)

Artificially, naturally

(Trilogy, Rikilt collected Maize, wheat, peanuts)

VIS, FLUO, NIR, Camera (not integrated

with VIS spectrometer)

M11

IFN (Rome, IT) Maize Absence of

contamination -

VIS, FLUO, Camera

(integrated) M15

ISPA (Bari, IT) Almond,

Maize ng/g

Artificially (Almond), naturally (maize)

VIS, FLUO, Camera

(integrated) M16

Rikilt (Wageningen, NL)

Almond, Maize

ng/g

Artificially (Almond), naturally (maize)

VIS, FLUO M18

Table 3: Mycotoxin (AFB1) contamination levels of artificially contaminated almond and

maize samples.

Sample Contamination level (μg/g)

Almond no

contamination ND*

Almond LOW 7.93

Almond MEDIUM 11.40

Almond HIGH 20.01


Page | 27


Maize no

contamination ND

Maize LOW 23.26

Maize MEDIUM 42.40

Maize HIGH 96.54

* ND: not detected by HPLC analysis (i.e. absence of contamination).

Table 4: Mycotoxin contamination levels of naturally contaminated maize samples from

certified supplier (Trilogy labs).

Sample Contamination level (ng/g)

Maize trilogy no

contamination ND*

Trilogy

contamination 18.8



Page | 28


Table 5: Relative mycotoxin contamination levels of naturally contaminated maize samples

(Rikilt measurement session).


Maize_Rikilt_31 4.47











Page | 29


Table 6: Relative mycotoxin contamination levels of naturally contaminated peanut

samples (Rikilt measurement session).


Peanut_Rikilt_46 17.75











Page | 30


Table 7: Relative mycotoxin contamination levels of naturally contaminated wheat samples

(Rikilt measurement session).


Wheat_Rikilt_56 ND*

Wheat_Rikilt_57 ND

Wheat_Rikilt_58 1.72

Wheat_Rikilt_59 ND

Wheat_Rikilt_60 ND

Wheat_Rikilt_61 ND



Page | 31


Table 8: Relative mycotoxin contamination levels of artificially contaminated almond

samples (M16 Bari measurement session and M18 Rikilt measurement session).


Almond 0 291

Almond 1 39.32

Almond 2 16.36

Almond 3 12.48

Almond 4 9.89

Almond 5 8.28

Almond 6 6.01

Almond 7 4.20

Almond 8 2.65


Page | 32


Table 9: Relative mycotoxin contamination levels of naturally contaminated maize samples

from Lugo town (M16 Bari measurement session and M18 Rikilt measurement session).


Maize Lugo 6 0.5

Maize Lugo 2 2.4

Maize Lugo 5 3.2

Maize Lugo 4 14.6

Maize Lugo 7 30.4

Maize Lugo 1 59.2


Page | 33


Table 10: Relative mycotoxin contamination levels of naturally contaminated maize

samples from Copparo town (M16 Bari measurement session and M18 Rikilt measurement

session).


Maize Copparo 8 3.9

Maize Copparo 9 11.8



3.2.1 Maize flour

3.2.1.1 Materials

Production of maize kernels contaminated with aflatoxins

Maize kernels were inoculated with a strain of Aspergillus flavus that produces aflatoxins B1 and

B2. In particular, maize kernels were subjected to a mild sterilization by immersion for 2 min in a

solution of 1% sodium hypochlorite with manual agitation. Subsequently, kernels were washed

with sterile water and then further surface-disinfected with 70% ethanol for 1 min. Kernels

were then transferred, under sterile conditions, in sterile flasks and portions of 1 kg of maize

were treated with 300 ml of a conidial suspension containing 3.3 × 105 conidia/ml of A. flavus.

The conidial suspension was obtained from A. flavus inoculated on plates of potato dextrose

agar (PDA) and grown for 4 days in the dark. Fungal conidia were collected from the surface of

plates by using a sterile spatula and sterile water. Appropriate dilutions with sterile water were

made to obtain a final concentration of 3.3 × 105 conidia/ml. After inoculation, flasks were

manually stirred to uniformly distribute the conidial suspension on kernels’ surface and closed

with a stopper of raw cotton and a sheet of aluminum foil to avoid excessive water

evaporation. The flasks were incubated for 4 days at 28°C in the dark. After incubation, maize

kernels were dried at 40°C for 48 h. Then, three portions of maize kernels (10 g each) were

randomly collected from each batch, content of flask, and analyzed for their aflatoxin content.


Page | 34


Determination of AFB1 and AFB2

Five grams of dry ground sample were extracted with 50 ml of extraction mixture of

acetone/water (85:15 v/v) by sonication for 30 min. The extract was filtered through a filter

paper and 0.1 ml was diluted with 6 ml of MeCN/H2O (30:70 v/v), stirred, filtered through a

0.45 μm PTFE and 0.1 ml was injected into the HPLC system (corresponding to 0.17 mg sample).

HPLC-FLD equipment and conditions

The HPLC-FLD analyses were performed with an Agilent 1260 consisting of a binary pump, an

autosampler with a 100 μl loop, a fluorescence detector fixed at 365 nm λex and 435 nm λem

and a thermostatic oven set at 30 °C. The column used was a 150 mm × 4.6 mm i.d., 3 μm, Luna

PFP (2) (pentafluorophenyl-propyl) with a 3 mm i.d., 0.45 μm pore size guard. The

chromatographic separation was performed in the isocratic condition using a mixture of

MeCN/H2O (30:70 v/v) at flow rate of 0.8 mL/min. A photochemical postcolumn derivatization

UVE was used to enhance the fluorescence of AFB1 and AFB2.

Granularity separation

When the separation was introduced, maize particles of different granularity dimensions

different granularities were separated with a sieve including different meshes (2 mm / 1 mm /

500 μm and 300 μm) and then considered for measurements separately. In subsequent

analysis, the spectra related to samples with different dimensions were analyzed separately. In

our notation, 1 mm refers to grained particles with dimensions > 500 μm but <1000 μm.

Certified reference material

Naturally contaminated (aflatoxin B1 18.8 µg/kg aflatoxin B2 0.9 µg/kg aflatoxin G1 2.4 µg/kg

aflatoxin G2 ND (µg/kg) deossinivalenol 2.6 mg/kg ochratoxin A 4.0 µg/kg T-2 Toxin 263.7

µg/kg HT-2 Toxin 523.3 µg/kg zearalenon 352.0 µg/kg fumonisin B1 28.3 mg/kg fumonisin B2

7.1 mg/kg fumonisin B3 1.7 mg/kg) and non-contaminated control maize flour samples

purchased from Trilogy Analytical laboratory were also used for feasibility tests.

Naturally contaminated maize

Naturally contaminated maize was received from different suppliers and at different

contamination levels (ranging from 0.5 to 87.3 ng/g), the contamination was quantified and the

maize was grained for spectral acquisition (see Tables 9 and 10).


Page | 35


3.2.1.2 Experimental design

The maize flour sample measurement set-up consists of a 100 mm or 60 mm diameter Petri

dish in which the maize flour is poured. Measurements at five different spatial positions over

the sample have been performed, and in each position five replicates have been acquired and

averaged. A preprocessing procedure including averaging and normalization steps was applied

during data processing. VIS spectrometer is positioned on top of the Petri dish and light is

shielded by a dark cover. Fluorescence measurements and VIS spectroscopy measurements

have been performed with this setting.

For the first set of measurements, the analyzed maize flour samples consisted of three different

levels of Aflatoxin B1 contamination (No cont= no aflatoxin contamination, Low cont= 23.3 μg/g,

High cont=96.5 μg/g) obtained by inoculation with the fungi A. flavus. Preliminary

measurements have also been performed on maize flour reference material from Trilogy Lab

contained in the Petri dish (18.8 μg/kg B1 aflatoxin). Also in this case, measurements at

different positions have been acquired. Together with these artificially contaminated samples,

measurements on naturally contaminated samples have been performed both in Rikilt lab

measurement session and in Bari (see Tables 3-10 for sample names and contamination levels).

During the measurement session in M11 at Rikilt, images have been acquired for all the

samples measured in that session. Also, during the measurement session in M16 in Bari, images

have been acquired by the camera for each measured sample.

3.2.1.3 Implementation of sensors (food science laboratory)

Fluorescence measurements and VIS spectroscopy measurements have been performed

according to the SOP described in section 3.1.2. With regard to the fluorescence

measurements, the integration time has been set at 100 or 200 ms, while a plastic black

reference has been used as a reference. In the case of VIS spectroscopy, the integration time

has been set at 250 μs, whereas a white paper has been used as white reference material.

Images have been acquired by using the Ximea Camera, according to the SOP described in

section 3.1.3, for all the measured samples.

3.2.1.4 Other experimental procedures

In our preliminary experiments, samples have been also characterized with standard techniques

like high-performance liquid chromatography (HPLC) for validation purposes.


Page | 36


3.2.2 Milk powder

Preliminary tests have not been yet performed but will be performed in the course of WP3. The

expected measurement setup will be similar to the ones used in maize and almond flour.

3.2.3 Paprika powder

Preliminary tests have not been yet performed but will be performed in the course of WP3. The

expected measurement setup will be similar to the ones used in maize and almond flour.

3.2.4 Tree nuts

3.2.4.1 Materials

Production of almonds contaminated with aflatoxins - First set: highly contaminated samples

(μg/g, i.e. ppm)

Shelled almonds (cs. Genco) were inoculated with a strain of Aspergillus flavus that produces

aflatoxins B1 and B2. In particular the almonds were subjected to a mild sterilization by

immersion for 2 min in a solution of 1% sodium hypochlorite with manual agitation.

Subsequently, almonds were washed with sterile water and then further surface-disinfected

with 70% ethanol for 1 min. Almonds were then transferred, under sterile conditions, in sterile

flasks and aliquots of 1 kg of almonds was treated with 300 ml of a conidial suspension

containing 3.3 × 105 conidia/mL of A. flavus. The conidial suspension was obtained from A.

flavus inoculated on plates of potato dextrose agar (PDA) and grown for 4 days in the dark.

Fungal conidia were collected from the surface of plates by using a sterile spatula and sterile

water. Appropriate dilutions with sterile water were made to obtain a final concentration of 3.3

× 105 conidia/ml. After inoculation, flasks were manually stirred to uniformly distribute the

conidial suspension on almond’s surface and closed with a stopper of raw cotton and a sheet of

aluminum foil to avoid excessive water evaporation. The flasks were incubated for 4 days at 28

°C in the dark. After incubation, almonds were dried at 40 °C for 48 h. Then three aliquots of

almonds (10 g each) were randomly collected from each batch, content of flask, and analyzed

for their aflatoxin content (see maize preparation).

Production of artificially contaminated almond samples - Second set: low contaminated

samples (ng/g, i.e. ppb)

For the second set of samples, contaminated sample #0 (291 ng / g, see Table 8) was prepared by mixing a culture of almonds contaminated with AFB1 (see paragraph above for preparation) and white almonds.


Page | 37


In particular, 3.12 g of minced almonds contaminated with AFB1 at a level of 35.5 μg / g were added to 496.87 g of ground white almonds. The almonds were homogenized by making slurry i.e. homogenizing the solid sample with water with a 1: 2 ratio (solid: liquid). In particular, one liter of water was added to 500 g of solid sample (3.12 g of contaminated almonds + 496.87 g of white almonds). The sample was homogenized with a T 25 digital ULTRA-TURRAX® for 5 min. Subsequently, the sample was frozen, lyophilized and analyzed using HPLC.

Samples # 1, # 2, # 3, # 4, # 5, # 6, # 7 and # 8 were prepared by making slurry by adding appropriate amounts of sample #0 to appropriate amounts of white almonds, exactly like sample #0 preparation. In particular, 80 ml of water were added to 80 g of each sample. The sample was homogenized as described previously, and was then frozen, lyophilized and analyzed using HPLC.

The levels of AFB1 contamination of samples # 1 through # 8 are presented in Table 8.


The experimental design for tree nuts is similar as for the maize flour with some minor

modifications. For almond, in the first set of measurements, four different levels of

contamination have been considered: no contamination, low contamination, medium

contamination and high contamination corresponding to 0, 20, 11, 4 and 7.9 μg/g, respectively.

In this case, contamination has also been obtained by inoculation with A. flavus. Grained

almond flour has been separated according to particles’ dimensions with the help of

mechanical sieves with sizes >2 mm / 2 mm / 1 mm and 500 μm. Depending on contamination

level, not all dimensions are available (contamination caused damage in the texture of the

sample matrix).

For the second set of measurements, artificially contaminated almond samples consisted of

nine different levels of Aflatoxin B1 contamination (see Table 8) obtained by inoculation with

the fungi A. flavus (Fig. 6). For each level of contamination, four samples have been prepared

and measured. Each sample has also been imaged with camera in white light mode (Fig. 7).

Also in this case, in M11 at Rikilt, images have been acquired for all the samples measured in

that session. Also, during the measurement session in M16 in Bari, images have been acquired

by the camera for each measured sample (see Fig. 7).


Page | 38


Figure 6: Second set of almond samples measured in Bari (April 2018) and in Wageningen (June 2018) (see Table

8).

Figure 7: Example of an almond sample image acquired with Ximea camera.


Fluorescence measurements and VIS spectroscopy measurements have been performed

according to the SOP described in section 3.1.2. In the case of fluorescence measurements, the

integration time has been set at 100 ms, while a plastic black reference has been used as a dark

reference. In the case of VIS spectroscopy, the integration time has been set at 400 μs, whereas

a plastic black reference and a white paper has been used as a dark reference and a white

reference material, respectively.


Page | 39



Determination of AFB1 and AFB2

Five grams of dry ground sample were extracted with 50 ml of extraction mixture of

acetone/water (85:15 v/v) by sonication for 30 min. The extract was filtered through a filter

paper and 0.1 ml was diluted with 6 ml of MeCN/H2O (30:70 v/v), stirred, filtered through a

0.45 μm PTFE and 0.1 ml was injected into the HPLC system (corresponding to 0.17 mg sample).

HPLC-FLD equipment and conditions

The HPLC-FLD analyses were performed with an Agilent 1260 consisting of a binary pump, an

autosampler with a 100 μl loop, a fluorescence detector fixed at 365 nm λex and 435 nm λem

and a thermostatic oven set at 30 °C. The column used was a 150 mm × 4.6 mm i.d., 3 μm, Luna

PFP (2) (pentafluorophenyl-propyl) with a 3 mm i.d., 0.45 μm pore size guard. The

chromatographic separation was performed in the isocratic condition using a mixture of

MeCN/H2O (30:70 v/v) at flow rate of 0.8 mL/min. A photochemical postcolumn derivatization

UVE was used to enhance the fluorescence of AFB1 and AFB2.

3.3 Detection of food spoilage and shelf-life prediction

3.3.1 Meat


The meat product evaluated under this sub-use case was minced pork, stored aerobically or in

modified atmosphere packaging (MAP) and at different temperatures. More specifically,

portions (100 g) of minced pork were stored under isothermal (4, 8 and 12°C) and dynamic

temperature (periodic temperature changes from 4 to 12°C) conditions, aerobically or under

MAP (80% O2-20% CO2) conditions for a maximum time period of 14 and 15 days, respectively.

The 100-g minced pork portions were shaped in patties, and duplicate patties were placed in

Styrofoam trays prior to storage under the aforementioned conditions in high-precision

(±0.5°C) programmable incubators (MIR-153, Sanyo Electric Co., Osaka, Japan). At regular time

intervals during storage, depending on the applied storage conditions (i.e. temperature and

atmosphere), duplicate minced pork patties were subjected to the following

analyses/measurements:

o Microbiological analyses, pH measurements and sensory evaluation (all samples)

o Multispectral image (MSI) acquisition (all samples)

o Fourier transform infrared (FTIR) spectroscopy measurements (all samples)


Page | 40


o NIR spectroscopy measurements (NIR spectrometer developed by IPMS)

(samples stored under MAP)

o VIS (all samples) and FLUO (samples stored under MAP) spectroscopy

measurements (VIS spectrometer developed by CNR)

o Image acquisition using the CMOS camera (samples stored under MAP)

A general schematic of the applied experimental analyses, applied in this (i.e. minced pork) as

well as in the rest of the sub-use cases falling under use case 2 and described subsequently in

this report, is provided in Figure 8. The MSI acquisition and the FTIR spectroscopy

measurements were performed as additional advanced spectroscopy methods, with the

ultimate goal of serving as point of reference for the presented PhasmaFOOD scanners in

relation to the spoilage status of the food samples analyzed. This is the case not only for this

specific food product (minced pork) but for all the products (sub-use cases) studied under use

case 2.

Two independent experiments (i.e. different time instances and different meat batches) were

conducted, and a total of 228 and 202 minced pork samples were analyzed during aerobic and

MAP storage, respectively.

Figure 8: Experimental analyses for the evaluation of the feasibility of the application of the PhasmaFOOD micro-sensors for the detection of food spoilage.


Page | 41



The NIR and VIS spectrometers evaluated in the PhasmaFOOD project were used in the minced

meat spoilage experiments following the standard operating procedure (SOP) described in

D3.1. The acquired NIR measurements were saved using Option 2 as described in section 3.1.1.

With reference to the VIS spectrometer and its use in the fluorescence modality, if saturation in

fluorescence signal is observed, then reduction of integration time to 100000 μs may be

required.


Microbiological analyses

Twenty five grams from each minced pork patty were transferred aseptically to a 400-ml sterile

stomacher bag (Seward Medical, London, United Kingdom) containing 225 ml of sterilized

Ringer’s solution (Lab M Limited, Lancashire, UK), and were homogenized for 60 sec (Lab

Blender 400, Seward Medical). For the enumeration of total mesophiles, appropriate serial

decimal dilutions in Ringer’s solution were surface plated on tryptic glucose yeast agar (Biolife,

Milan, Italy), and colonies were counted after incubation of plates at 25°C for 72 h.

Furthermore, at some of the sampling intervals, the meat samples also were tested for the

determination of the populations of Pseudomonas spp., Brochothrix thermosphacta, lactic acid

bacteria (LAB), bacteria of the family Enterobacteriaceae and molds and yeasts. Pseudomonas

spp. were enumerated by surface plating on Pseudomonas agar base supplemented with CFC

(Cephalothin, Fucidin, Cetrimide) Selective Supplement (Lab M Limited) after incubation of

plates at 25°C for 48 h. Br. thermosphacta was enumerated by surface plating on streptomycin-

thallous acetate-actidione (STAA) agar (Biolife) after incubation of plates at 25°C for 48 h. LAB

were enumerated by pour plating in de Man Rogosa and Sharpe (MRS) agar (Biolife) after

incubation of plates at 30°C for 72 h. Bacteria belonging to the Enterobacteriaceae family were

enumerated by pour plating in violet red bile glucose (VRBG) agar (Biolife) after incubation of

plates 37°C for 24 h. Finally, molds and yeasts were enumerated by surface plating on Rose

Bengal Chloramphenicol (RBC) agar (Lab M Limited) after incubation of plates at 25°C for a total

of 5 days. The obtained microbiological data were converted to log (colony forming units) per

gram of minced meat (log CFU/g). Upon completion of the microbiological analyses, the pH

values of the meat samples also were measured using a digital pH meter (RL150, Russell pH,

Cork, Ireland) with a glass electrode (Metrohm AG, Herisau, Switzerland). Finally, each meat

sample was evaluated with regard to its sensory attributes of appearance and odor. Specifically,

each attribute was scored on a three-point hedonic scale where: 1=fresh; 2=marginal; and

3=unacceptable. A score of 1.5 was used to characterize a sample as semi-fresh and was

regarded as the first indication of change from that of typical fresh meat. Scores above 2

rendered the product spoiled and were regarded as indicative of the end of the product’s shelf-

life.


Page | 42


Multispectral image (MSI) acquisition

Images from each minced meat sample were acquired using the VideometerLab system,

originally developed by the Technical University of Denmark [31] and commercialized by

“Videometer A/S” (http://www.videometer.com). This instrument, which has been described in

detail elsewhere [26, 32], acquires multispectral images in 18 different, non-uniformly

distributed wavelengths ranging from UV (405 nm) to short wave NIR (970 nm). Prior to image

acquisition, the system was subjected to a light set up procedure known as “autolight” and

calibrated radiometrically and geometrically as previously described [26]. Portions from each

minced meat sample were placed in a Petri dish which was then was placed (without the plate’s

lid) inside an Ulbricht sphere in which the camera is top-mounted, and the corresponding

multispectral image of the product’s surface was taken.

FTIR spectroscopy

FTIR spectral data were collected using a ZnSe 45° HATR (Horizontal Attenuated Total

Reflectance) crystal (PIKE Technologies, Madison, Wisconsin, United States), and an FT/IR-6200

JASCO spectrometer (Jasco Corp., Tokyo, Japan) equipped with a standard sample chamber, a

TGS detector and a Ge/KBr beamsplitter. A small portion from each minced meat sample was

transferred in the crystal plate and covered with a small piece of aluminum foil, and then

pressed with a gripper in order for the best possible contact with the crystal to be achieved.

The crystal used has a refractive index of 2.4 and a depth of penetration of 2.0 μm at 1000 cm-1.

Using the Spectra Manager™ CFR software version 2 (Jasco Corp.), spectra were collected over

the wavenumber range of 4000 to 400 cm-1, by accumulating 100 scans with a resolution of 4

cm-1 and a total integration time of 2 min. Prior to the measurements of the tested samples,

reference spectra were acquired using the cleaned blank (no added food sample) crystal. After

each measurement, the crystal’s surface was cleaned, first with detergent and distilled water

and then with analytical grade acetone, and dried using lint-free tissue. The FTIR spectra that

were ultimately used in further analyses were in the approximate wavenumber range of 3100

to 2700 and 1800 to 900 cm-1.

3.3.2 Fish


The experimental procedure designed and performed under this section aimed at the

monitoring and evaluation of the spoilage of gilthead sea bream (Sparus aurata) during aerobic

or MAP (30% O2, 40% CO2, 30% N2) storage under different isothermal conditions. For this

purpose, aquacultured whole ungutted fish, within two days from harvest, were packaged in

Styrofoam trays wrapped by cling film (aerobic storage) or placed in plastic bags (MAP storage)

http://www.videometer.com/


Page | 43


and stored at 0, 4 and 8°C in high-precision (±0.5°C) programmable incubators (MIR-153, Sanyo

Electric Co.) for a maximum time period of 11 and 19 days, respectively. At regular time

intervals during storage, depending on the applied storage conditions (i.e. temperature and

atmosphere), duplicate fish samples (originating from different fish) were subjected to the

following analyses/measurements:


o MSI acquisition (all samples)

o FTIR spectroscopy measurements (all samples)

o NIR spectroscopy measurements (all samples)

o VIS and FLUO spectroscopy measurements (samples stored under MAP)

The applied experimental analyses were performed according to the general schematic

illustrated in Figure 8. After analysis of spectral data acquired on both the skin and the flesh

sides of the fish samples (as described in detail in D3.1 for aerobic storage of fish), it was

decided that data corresponding only to fish skin will be evaluated during subsequent

experiments under MAP, since only this side of the samples appears to have considerable

correlation with the microbiological spoilage of whole ungutted fish.

Two independent experiments (i.e. different time instances and different fish batches) for each

storage condition were conducted, and a total of 158 and 183 fish were analyzed during aerobic

and MAP storage, respectively.


The NIR and VIS spectrometers evaluated in the PhasmaFOOD project were used in the fish

spoilage experiments following the SOPs described in D3.1. The acquired NIR measurements

were saved using Option 2 as described in section 3.1.1. With particular reference to the VIS

spectrometer and the measurements acquired in fluorescence mode, in order to avoid to the

greatest possible extent signal saturation issues, the integration time used was 100000 μs.



Ten grams from each fish were transferred aseptically to a 400-ml sterile stomacher bag

(Seward Medical) containing 90 ml of sterilized peptone saline diluent (0.1% w/v peptone,

0.85% w/v sodium chloride), and were homogenized for 60 sec. For the enumeration of total

mesophiles, appropriate serial decimal dilutions in peptone saline diluent were surface plated

on tryptic glucose yeast agar, and colonies were counted after incubation of plates at 30°C for

72 h. Furthermore, at some of the sampling intervals, the fish samples also were tested for the


Page | 44


determination of the populations of Pseudomonas spp., Br. thermosphacta, LAB, bacteria of the

family Enterobacteriaceae and molds and yeasts. Pseudomonas spp. were enumerated by

surface plating on Pseudomonas agar base supplemented with CFC Selective Supplement after

incubation of plates at 25°C for 48 h. Br. thermosphacta was enumerated by surface plating on

STAA agar after incubation of plates at 25°C 48 h. LAB were enumerated by pour plating in MRS

agar after incubation of plates at 30°C for 5 days. Bacteria of the family Enterobacteriaceae

were enumerated by pour plating in VRBG agar after incubation of plates at 37°C for 24 h.

Finally, molds and yeasts were enumerated by surface plating on RBC agar after incubation of

plates at 25°C for 5 days. The obtained microbiological data were converted to log CFU/g. Upon

completion of the microbiological analyses, the pH values of the fish samples also were

measured using a digital pH meter with a glass electrode. Finally, each fish sample was

evaluated with regard to its sensory attributes of appearance and odor. Specifically, each

attribute was scored on a three-point hedonic scale where: 1=fresh; 2=marginal; and

3=unacceptable. A score of 1.5 was used to characterize a sample as semi-fresh and was

regarded as the first indication of change from that of typical fresh fish (i.e. clear eyes, odor

slightly changed, but still acceptable by the consumer). Scores above 2 rendered the product

spoiled and were regarded as indicative of the end of the product’s shelf-life.

Multispectral image (MSI) acquisition and FTIR spectroscopy

Multispectral imaging and FTIR spectral data from each fish sample were collected by using a

cut fish portion, according to the procedures described previously (section 3.3.1.3).

3.3.3 Fruit and vegetables


The experimental procedures designed and conducted under this section aimed at the

monitoring and evaluation of the spoilage of (i) rocket salad (Eruca sativa), (ii) baby spinach

(Spinacea oleracea) and (iii) fresh-cut pineapple (Ananas comosus) during aerobic storage

under different temperature conditions. All these three food products were washed and ready-

to-eat fresh produce commodities provided by a local manufacturer within one day from

production, and were used in the experiments in their original commercial packaging (i.e.

plastic bags for rocket/spinach and plastic cups with lid for pineapple). The aforementioned

fruit and vegetables were stored under isothermal conditions (i.e. 4, 8 and 12°C), as well as at

dynamic storage conditions with periodic temperature changes ranging from 4 to 12°C (i.e. 8 h

at 4°C, 8 h at 8°C and 8 h at 12°C). At regular time intervals during storage, depending on the

storage temperature, duplicate samples (originating from different packages) were subjected to

the following analyses/measurements:


Page | 45



o MSI acquisition (all samples)

o FTIR spectroscopy measurements (all samples)

o VIS (all samples) and FLUO (pineapple samples) spectroscopy measurements

o NIR spectroscopy measurements (all vegetable samples and some pineapple

samples)

o Image acquisition using the CMOS camera (baby spinach samples)

The applied experimental analyses were performed according to the general schematic

illustrated in Figure 8. As mentioned previously for the sub-use cases of meat and fish,

measurements taken with a vibrational spectroscopy instrument (i.e. FTIR spectrometer) and

surface chemistry (e.g., MSI) data acquisition were performed as additional advanced

metabolomics/fingerprinting methods, with the ultimate goal of serving as point of reference

for the presented PhasmaFOOD scanners in relation to the spoilage status of the food samples

analyzed. Two and three independent experiments (i.e. different time instances and different

product batches) were conducted for the vegetable (rocket, spinach) and fruit (pineapple)

products, respectively. A total of 232 rocket samples, 224 baby spinach samples and 318

pineapple samples were analyzed in the context of the abovementioned experiments.


The NIR and VIS spectrometers evaluated in the PhasmaFOOD project were used in the fruit

and vegetables spoilage experiments following the procedures described in D3.1. With

particular reference to the VIS spectrometer and the measurements acquired in fluorescence

mode (i.e. in the case of pineapple samples), in order to avoid to the greatest possible extent

signal saturation issues that arose at some instances, the integration time used in the third

experimental replicate was 100000 μs.



Twenty-five grams from each produce commodity sample (i.e. commercial package of fruit or

vegetables) were transferred aseptically to a 400-ml sterile stomacher bag containing 225 ml of

saline diluent, and were homogenized for 60 sec. For the enumeration of total mesophiles,

appropriate serial decimal dilutions in saline diluent were surface plated on tryptic glucose

yeast agar, and colonies were counted after incubation of plates at 25°C for 72 h. In addition,

Pseudomonas spp. populations and total populations of molds and yeasts throughout storage,

as well as populations of LAB and Enterobacteriaceae at selected time intervals during storage,

were determined according to the procedures described in detail in the sections 3.3.1.3 and

3.3.2.3. The obtained microbiological data were converted to log CFU/g. Upon completion of


Page | 46


the microbiological analyses, the pH values of the fruit and vegetable samples also were

measured using a digital pH meter with a glass electrode. Finally, each produce commodity’s

sample was evaluated with regard to its sensory attributes of appearance (color) and odor,

using the procedures described previously (sections 3.3.1.3 and 3.3.2.3).

Multispectral image (MSI) acquisition and FTIR spectroscopy

Multispectral imaging and FTIR spectral data were collected for the vegetable salad samples by

using an appropriate amount of whole and cut leaves, respectively, according to the procedures

described previously (section 3.3.1.3). With regard to the tested fruit (i.e. pineapple) samples,

portions of pineapple slices and small cut portions of the fruit were used for the purpose of

image acquisition and FTIR spectra collection, respectively.

3.3.3.4 Pilot experiments

The available NIR spectrometer was furthermore tested on a wider range of fruit, with two

purposes: 1) these measurements served to assess the amount of light required for the

PhasmaFOOD sensing system. They were carried out alongside sensor development; 2) the fruit

tested included avocado, mango, papaya, banana, strawberries and other fruit that are

considered sensitive from the point of food quality deterioration, e.g. where the state of

“perfect ripeness” occurs only within a very short time window or where ripeness reacts very

sensitive to storage conditions. These measurements are not fully evaluated yet and therefore,

not shown. They can be used to assess whether it is worthwhile to run a full trial on a certain

type of fruit, in addition to pineapple.

3.4 Detection of food fraud

3.4.1 Milk powder


Thirty-two skimmed milk powders (SMP) were obtained from routine controls (residue

samples) which had passed the routine controls as described in EC regulation 273/2008 on

public intervention for butter and SMP [33]. Milk powder adulteration experiments were

divided in two groups: (I) safety related experiments using nitrogen enhancers and (II) non-

hazardous low-value fillers.

For the safety related issues, the following nitrogen enhancers were considered: melamine,

ammonium chloride, ammonium nitrate and urea. All chemicals were purchased in pure or


Page | 47


technical grades of purity. Nitrogen enhancers were dry blended1 in a full factorial design in

concentrations of 0.1 to 5.0 % (w/w), which results in an apparent raise of protein

concentration of approximately 0.1 to 2.0 % (w/w). For each nitrogen enhancer, nine blends

were prepared in the indicated concentration range. In total, 36 blends were prepared. For

each blend, a SMP was used from a different batch.

For the non-hazardous fillers, whey protein isolates (WPI, 6 commercial brands), plant protein

isolates (PaPI, 5 commercial brands), pea protein isolates (PePI, 4 commercial brands), soy

protein isolates (SPI, 3 commercial brands), buttermilk powder (BP, reference standard), starch

(S), maltodextrin (MD), glucose (G), lactose (L) and fructose (F) were considered. All sugars were

of laboratory synthesis grade. Again, dry blending was applied in a similar design as for the

safety related nitrogen enhancer, but now at concentrations of either 50 SMP / 50 filler or 25

SMP / 25 filler (% w/w) leading to a total of 36 blends.


The FLUO and VIS measurements were performed according to the SOP in section 3.1. SMPs

were deposited in plastic Petri dishes (diameter 9.0 cm) and measurements were performed by

placing the sensor casing on the outer rims of the Petri dishes. The measuring setup and

samples were shielded for environmental light by applying a black plastic bag around the setup.

The NIR measurements were performed according to the SOP in section 3.1 using the solids

probe which was connected to the device by an optical fiber. The probe was applied directly in

the sample flask which contained approximately 5 - 7 cm SMP powder.


All SMPs were analyzed for dry matter, fat content, protein content, acid whey content,

buttermilk content and carbohydrate content according to the accompanying ISO methods

stated in EC regulation 273/2008 [33, 34].

1 It is imperative to realize that SMP can be measured in dry form or in reconstituted form. Karunathilaka et al.

(2017) utilized Raman spectroscopy to classify non-adulterated milk powder and melamine adulterated milk powders, made using dry or wet blending methods. Utilizing PCA to analyze their spectra, the authors noted that dry blended adulterated samples were more difficult to differentiate from authentic samples, and that replicates were not tightly clustered [35]. Possible differences arising from the use of dry blending versus wet blending have also previously been mentioned by Capuano et al. (2015) [36]. In this sub-use case, dry blending was used for reasons of practicality in sample preparation.


Page | 48


3.4.2 Meat


The fraud incident studied under this section, covering part of the activities outlined in T3.3 of

WP3, is minced meat adulteration. More specifically, aiming at evaluating the efficacy of the

PhasmaFOOD sensors in detecting the fraudulent substitution of a specific type of raw minced

meat with other, two scenarios of minced meat adulteration were studied:

(I) Adulteration of chicken and pork

(II) Adulteration of beef and offal

In the first series of adulteration experiments, the objective was to assess the potential of the

PhasmaFOOD sensors, supported by advanced statistical approaches, for the detection of

minced chicken intentionally adulterated with minced pork and vice versa. For this purpose,

fresh pork cuts (leg) and chicken breast were obtained from different butcher shops and were

ground separately one at a time, using a domestic meat-mincing machine in the Laboratory of

Microbiology and Biotechnology of Foods (Department of Food Science and Human Nutrition,

AUA). In the second series of adulteration experiments, aiming at evaluating the potential of

the PhasmaFOOD sensors for the detection of fraudulent adulteration of minced beef with

offal, similar procedures were followed for the preparation of minced meat derived from fresh

beef cuts (leg) and bovine hearts. In both series of experiments (adulteration scenarios I and II),

different levels of adulteration, ranging from 25 to 75%, were achieved by mixing the

appropriate amount of each type of meat under conditions simulating industrial processing. A

total of five different levels of adulteration were considered for each adulteration scenario,

consisting of three categories of mixed meat and two categories of pure meat. From each level

of adulteration, six different portions (ca. 70 g) were placed in Petri dishes and the following

measurements were taken:

o MSI acquisition (i.e. snapshots using the VideometerLab system)

o NIR spectroscopy measurements

o VIS and FLUO spectroscopy measurements

For each type of minced meat adulteration (adulteration scenario: chicken/pork, beef/offal)

and for each adulteration level within a type, each Petri dish was considered as a replicate in

the experiment (i.e. 5 adulteration levels × 6 samples in total per experiment). Each

experimental procedure was repeated four times (i.e. four independent experiments

corresponding to distinct meat batches), resulting in 120 meat samples being studied in total

for each one of the two aforementioned adulteration scenarios.


Page | 49



NIR, VIS and FLUO spectroscopy measurements were performed according to the procedures

described previously (section 3.3.1.2).


Similarly to what was the case for the use case 2 experiments, MSI data acquisition served as

point of reference for the presented PhasmaFOOD scanners in relation to meat adulteration

[26, 28].

3.4.3 Alcoholic beverages


Approximately fifty distilled spirit samples are obtained from various liquor shops in The

Netherlands from various categories: bourbon, Dutch gin (light, young and old), gin, grain wine,

grappa, whiskey and vodka. This sub-use case is divided in assessing (I) dilution (i.e. prediction

of ethanol content), (II) presence of technical alcohols (i.e. methanol) and (III) counterfeit

products. During this first feasibility test the PhasmaFOOD sensors will be tested for dilution

(range 15 – 50 % alcohol by volume (abv)) and presence of methanol (range 1 – 20 % abv), in a

factorial sampling design. The production of counterfeit samples for ‘white’ spirits like vodka,

(Dutch) gin by water/ethanol mixtures will be attempted when acceptable results for dilution

and technical alcohols are obtained.


Preliminary tests have been performed using the NIR sensor according to the SOP described in

section 3.1. Tests were performed in the transmission modus (cuvette holder and

complementary fiber optics), using a 1-cm plastic standard issue cuvette. The FLUO and VIS

measurements were not performed as the device was not suitable for performing

measurements on liquids.


The alcoholic strength and determination of methanol will be executed accor ding to EC

regulation 2870/2000 [37] on the reference methods for the analysis of spirit drinks.


Page | 50


3.4.4 Edible oils

Pilot experimental results using the NIR spectrometer and mixtures of extra virgin olive oil and

sunflower oil were reported in D3.1. Feasibility will be further assessed and reported in the

further course of WP3, in D3.7.


Page | 51


4 Data analysis

The data processing could be divided in two main parts. The first part is related to the

acquisition of the measurements from different sensor systems (NIR spectrometer, VIS

spectrometer and CMOS camera) while the second is mainly composed of the following steps:

1. Data Preprocessing: data compression (for the sensor required, noise removal, data

normalization for de-correlation and to have the same range of values for each of the sensors

(this can guarantee stable convergence of the downstream developed models).

2. Feature selection: identification for each case study (exploiting the information that we

have obtained from the preliminary measurements) of the most significant features for each

sensor device. In this way, it is not necessary to declare a priori which kind of sensors will be

used for the specific case study.

3. Classification/identification/regression: In this step, using the most significant features

as input for ad hoc use case model will be applied to the data in order to have as output a

sample prediction.

Since D3.3, which coincides in time in the course of the project with D3.2, is particularly

devoted to the description of detection algorithms, detailed description of the chemometric

algorithms and data fusion strategies (pertinent to the smart data correlation activities outlined

in T3.4 of WP3) is provided in the former deliverable’s report. In this context, only a brief

presentation of the data analysis procedures is provided in this report.

4.1 Mycotoxin detection

At the present stage of tests, detection of mycotoxins (particularly aflatoxin B1) has relied

mainly on fluorescence emission spectra. Visible reflectance data have been recorded and will

be considered in multivariate analysis. The data analysis is mainly based on the three steps

previously cited. In the preprocessing, a normalization procedure has been applied to each

single spectrum. In particular, the normalization with dark and reference samples is applied for

VIS spectra while only the dark normalization is considered for FLUO spectra. After that, the

processed spectra are arranged in a matrix and used as input for partial least squares regression

for a regression task aimed at predicting mycotoxins content and for Linear Discriminant

Analysis for the classification task aimed at discriminating contaminated and non-contaminated

samples. The validation procedure has been performed by means of k-fold cross validation

procedure.


Page | 52


4.1.1 Maize flour

With regard to the VIS spectral data, these consisted of 500 measurements at 288 wavelengths.

Ten different concentrations of aflatoxins were considered ranging from 0.96 to 5.47 ppb (i.e.

0.96, 1.27, 4.57, 4.70, 4.88, 4.97, 5.01, 5.11, 5.21 and 5.47 ppb). For each maize sample, five

different measurements (zones) were performed and ten repetitions were carried out. This

strategy is adopted to evaluate in the measurement the sample and instrument variability.

Regarding data preprocessing, each single spectrum was normalized using reference and dark

samples that were acquired prior to each sample measurement. The first part of data analysis is

related to the regression task. In this context, partial least squares (PLS) regression was used to

build the regression model. The number of latent variables (LVs) was determined by an internal

cross validation (10-fold) procedure using the training data as shown in Figure 9. Three LVs are

selected between the cross validation procedures with an error in cross validation equal to 0.5

ppb. In order to improve the reliability of the results, half of the samples (alternated) were used

for training the PLS model and the remaining samples for the purpose of model testing.

Figure 9: The mean squared error (MSE) values vs. the number of Latent Variable (LV) components.

Regarding the FLUO spectral data, the same data analysis approach as for the VIS data was

applied. The dataset consisted of 150 measurements at 288 wavelengths. The samples were the

same as the ones used in the VIS measurements, with the only differences lying in the number

of measured zones for each sample (i.e. three) and the number of replicates (i.e. three). As

described above, each single spectrum was normalized using a dark sample, acquired prior to

each sample measurement, and applying baseline normalization. Also in this case, half of the

samples were considered for training the PLS model and the remaining samples for external


Page | 53


validation purposes. The number of LVs was determined by an internal cross validation (10-

fold) procedure using the training data, as previously described.

The analyses of the data acquired during the last two measurement campaigns (M16 and M18)

are in progress and will be reported in further course of WP3.

4.1.2 Tree nuts

A data analysis strategy, similar to the one applied to the maize samples, has been performed

on the almond spectral dataset. The preliminary results of the data described in Table 8 and

measured in Bari at M16 are reported subsequently in the document of the present report (see

section 5.1.4).

4.2 Food spoilage detection and shelf-life prediction

4.2.1 Meat

The analysis of data derived from the minced meat spoilage experiments is in progress and will

be reported in full in the further course of WP3. The only data analyzed up to this point are the

ones corresponding to aerobic storage of minced pork and generated from the MSI and FTIR

spectroscopy sensors. In brief, the collected MSI and FTIR data were subjected to pre-

processing, i.e. smoothing based on SNV transformation and the Savitzky Golay algorithm,

respectively. Partial least squares (PLS) regression was used to establish the correlation

between imaging/spectral data and microbial counts, with the former constituting the input

and the latter the output variables in the PLS regression models. The models were calibrated

and validated with the data collected from the studied isothermal (170 samples) and dynamic

temperature (58 samples) conditions, respectively.

4.2.2 Fish

The collected data and data analysis procedures corresponding to the first fish fillets

experiment under aerobic storage conditions (pilot data acquisition), for all available, at the

time, sensors have been presented in the report of D3.1. The analysis of data corresponding to

the subsequent (i.e. after M9) fish spoilage experiments (i.e. under modified atmosphere

packaging (MAP) conditions), and involving both NIR and VIS spectral measurements, is in

progress and will be reported in the further course of WP3.


Page | 54



In the case of the data derived from the experiments on rocket salad, the processing pipeline

consisted of feature selection (specific wavelengths) on the basis of random forest (RF)

regression ensemble [38] followed by PLS regression for microbial contamination and shelf-life

estimation using the features (wavelengths/wavenumbers) selected by the RF ensemble. The

data analysis workflow, which is common more or less for all tested sensors, is described in

detail in D3.3.

The analysis of the data derived from the experiments on the spoilage of baby spinach and

pineapple is in progress and will be reported in the further course of WP3.

4.3 Food fraud detection

4.3.1 Milk powder

The data analysis of SMPs was performed on FLUO, VIS and NIR data. Due to time constraints,

the authentic SMP samples were measured in triplicate, but the adulterated blends were

measured only once, and need to be repeated as soon as the sensors become available again.

For data analysis, a chemometric classification strategy was used which will enable the

PhasmaFOOD sensor to detect a broad range of adulterants, even for compounds that the

model was not trained for. This technique is called one-class classification (OCC), where models

are trained to recognize the normal (spectral) variation for an authentic product, and flag

samples as (potentially) adulterated when observed spectra are outside this normal variation.

Data fusion was performed by applying SIMCA on the 1152 (one-) class distances resulting from

the OCC calculation. A detailed description of this strategy is given in D3.3.

4.3.2 Meat

The analysis of the data derived from the experiments on minced meat adulteration is in

progress and will be reported in the further course of WP3.


This dataset is not complete yet, due to the configuration of the FLUO and VIS sensor.

Feasibility of this sensor using OCC will be assessed in the final version of this deliverable, in the

report of D3.7.


Page | 55


4.3.4 Edible oils

This dataset is not complete yet, due to the configuration of the FLUO and VIS sensor.

Feasibility of this sensor using OCC will be assessed in the final version of this deliverable, in the

report of D3.7.


Page | 56


5 Results and discussion

5.1 Feasibility of micro-sensor application for detection of mycotoxins

5.1.1 Maize flour

The results of the preliminary tests that were performed on maize samples (including reference

samples) have been presented in the D3.1 report. With regard to the VIS reflectance

measurements, some representative maize samples spectra are illustrated in Figure 10.

Figure 10: Typical visible reflectance spectra corresponding to maize flour samples.

Some of the VIS spectra considered in the analysis, after the applied normalization procedure,

are shown in Figure 11, in particular the ones referring to the samples described in Table 5. It is

clearly evident that there is a sort of correlation between the acquired spectrum and the

relative aflatoxins concentration in the maize samples.


Page | 57


Figure 11: Examples of the acquired VIS spectra of maize flour samples after the applied data normalization procedure. The values in the legend are in ppb (ng/g).

The predicted versus measured aflatoxins concentration plot is shown in Figure 12. It is

interesting to remark that, although the prediction is not perfect (should lie on the bisector line

of the quadrant), the separation between low concentration (<4ppb) and high concentration

(>4ppb) is clear.

Figure 12: Predicted versus measured plot for the partial least squares (PLS) model based on VIS spectral data.

Moving to the collected FLUO data, reflectance spectra acquired prior to and after the applied normalization procedure are illustrated in Figure 13.


Page | 58


Figure 13: Raw and normalized fluorescence spectra corresponding to maize flour samples.

The predicted versus measured aflatoxins concentration plot, along with the values of the root mean square error (RMSE) of validation and testing, is shown in Figure 14.

Figure 14: Predicted versus measured plot for the partial least squares (PLS) model based on FLUO spectral data.

Additional tests and analysis have been performed, also considering the NIR spectra data but

without obtaining any interesting results. Plots of the raw NIR spectra and the first and second

derivatives for the maize measurements are shown in Figure 15.


Page | 59


The analysis of the data acquired during the last two measurement campaigns (M16 and M18)

is in progress and will be reported in further course of WP3.

Figure 15: Raw and transformed (first and second derivative after smoothing) NIR data corresponding to maize

flour samples.

5.1.2 Milk powder

Preliminary tests have not been performed yet.

5.1.3 Paprika powder

Preliminary tests have not been performed yet.

5.1.4 Tree nuts

In this section, preliminary data analysis results with regard to the almond samples

measurements, taken in the context of a novel measurement campaign (Bari, Italy), are

presented and briefly discussed. In this experimental campaign, all the ten contamination

levels, previously presented (Table 8) and also depicted in Table 11, were divided in four

replicates.


Page | 60


Table 11: Mycotoxin contamination levels of almond samples and their potential

assignment in classes.

Contamination

level (ppb)

Potential class division

(Good/No Good)

0 Good

2 Good

4 Good

6 No Good

8 No Good

12 No Good

16 No Good

40 No Good

291 No Good

The analysis of the generated fluorescence spectral data has been initiated, and the preliminary

results obtained for data collected during the first day for all samples are presented herein. The

spectra collected for sample #1 and for all contamination levels are illustrated in Figure 16. It is

interesting to note that as each spectrum was normalized with respect to the peak in order to

reduce the UV Led variability, there was no linear relationship between the spectrum and the

progressively increasing contamination levels (Fig. 16).


Page | 61


Figure 16: Representative fluorescence spectra of an almond sample normalized with respect to the peak.

Aiming at predicting the mycotoxin contamination level, a regression model was then

developed by means of the partial least squares (PLS) technique. In Figure 17 is shown the

evolution of the RMSECV as function of the number of Latent Variables (LV) considered in the

PLS model using k-fold cross validation procedure. The optimal number of LV is 9, obtaining a

good agreement between the real (observed) and predicted sample contamination, as

illustrated in Figure 18.

Figure 17: Evolution of the root mean square error of cross validation (RMSECV) as a function of the partial least

squares (PLS) components.


Page | 62


Figure 18: Predicted versus measured mycotoxin contamination obtained with partial least squares (PLS)

regression model using as input the FLUO spectral data.

The aforementioned results have been obtained considering the dataset of the first day in the

cross validation procedure. It is interesting to note that the most informative and interesting

part for the model is related to the contamination range of 0-40 ng/g. For this reason, a model

was then built excluding the sample at the highest concentration, and the resulting estimated

error was lower than that estimated in the previous case. Moreover, the number of LV was

reduced to 4, suggesting also a better capability of the PLS model to find the portion of the

spectrum mainly correlated with mycotoxin contamination (Figs. 19 and 20).


Page | 63


Figure 19: Evolution of the root mean square error of cross validation (RMSECV) as a function of the partial least

squares (PLS) components, excluding the sample at the highest concentration.

Figure 20: Predicted versus measured mycotoxin contamination obtained with partial least squares (PLS)

regression model using as input the FLUO spectral data excluding the highest contamination samples.


Page | 64


The last part of the preliminary analysis of the almond dataset is related to the evaluation of

the model performance with the data of the same day with an external dataset (Fig. 21).

Figure 21: Predicted versus measured plot for the partial least squares (PLS) regression model for training (left

panel) and test (right panel) phase. Testing is performed using external validation.

According to these last results, obtained considering all the contamination levels of one of the

four samples as external test, a sufficient generalizability of the results in the same day was

demonstrated. Nonetheless, a deeper investigation should be performed in order to find the

best model to reduce the estimation of mycotoxin contamination to lower levels. In this

context, data analysis will be continued with particular attention to the generalizability of the

regression and classification models in different days, also paying attention to the mycotoxin

contamination detection limits.

From this preliminary data analysis, it is interesting to note that there is a clear separation

between samples with contamination levels less than 60 ppb and higher than 60 ppb. So, at this

stage of the project and with the available integrated sensor system, we are confident that we

are able to detect a contamination level of 200 ppb with a recognition rate higher that 85%.

The analysis of the data acquired during the last measurement campaign (M18) is in progress

and will be reported in further course of WP3.

Train Test


Page | 65


5.2 Feasibility of micro-sensor application for detection of spoilage

5.2.1 Meat

5.2.1.1 Microbiological data

The results regarding the evolution of microbial populations (total mesophiles) during aerobic

and MAP storage of minced pork under different temperature conditions are presented in

Figure 22. As expected, the higher the storage temperature the higher the microbial growth

rate, and hence, the faster the spoilage of minced pork, both from a microbiological (Fig. 16)

and a sensorial (data not shown) perspective. The microbial growth appeared to be similar at

the constant temperature of 8°C and the dynamic (i.e. periodically changing) temperature

conditions under both studied atmospheres. Moreover, based on the microbiological data

derived for the specific microbial groups that were examined (data not shown), it appears that

Pseudomonas spp. and Br. thermosphacta constitute the dominant spoilage microflora (i.e.

specific spoilage organisms) of minced during aerobic and MAP storage, respectively. Finally,

although it was expected that storage under MAP would retard microbial growth (and thus,

enhance the shelf-life of minced pork), such an observation was not made herein. The most

likely explanation for this is the considerably higher initial microbial load of the minced pork

meat used in the MAP storage experiments compared to the one used in the aerobic storage

experiments.


Page | 66


Figure 22: Mean microbial populations during aerobic (A) and MAP (B) storage of minced pork at different isothermal (4, 8 and 12°C) and dynamic (8 h at 4°C, 8 h at 8°C and 8 h at 12°C) temperature conditions.

The analysis of data derived from the minced meat spoilage experiments is in progress and will

be reported in full in the further course of WP3. The only data analyzed up to this point are the

ones corresponding to aerobic storage of minced pork and generated from the MSI and FTIR

spectroscopy sensors.

5.2.1.2 Multispectral imaging (MSI) sensor (VideometerLab system)

MSI appeared to be promising in the quantitative evaluation of the microbiological spoilage of

minced pork. The coefficient of determination (R2) and the root mean square error (RMSE) for

the validation (prediction) of the PLS regression model based on the MSI data (Fig. 23) were

0.749 and 1.12, respectively.


Page | 67


Figure 23: Comparison between the observed and estimated by the PLS regression model total mesophilic microbial populations based on the MSI data of minced pork stored under aerobic conditions for the training (open symbols, 170 samples) and the testing (solid symbols, 58 samples) data sets (solid line: the ideal y=x line; dashed lines: the ± 1 log unit area).

5.2.1.3 FTIR spectroscopy

FTIR spectroscopy also appeared to be a promising rapid technique for the quantitative

monitoring of the microbiological spoilage of minced pork, with the developed model exhibiting

an even better performance of that based on MSI data. Specifically, the values of the coefficient

of determination (R2) and the root mean square error (RMSE) for the validation (prediction) of

the PLS regression model based on the FTIR data (Fig. 24) were 0.834 and 0.91, respectively.


Page | 68


Figure 24: Comparison between the observed and estimated by the PLS regression model total mesophilic microbial populations based on the FTIR data of minced pork stored under aerobic conditions for the training (open symbols, 170 samples) and the testing (solid symbols, 58 samples) data sets (solid line: the ideal y=x line; dashed lines: the ± 1 log unit area).

The analysis of the rest of the generated data (i.e. data derived from the PhasmaFOOD sensors

for aerobic storage and all data corresponding to MAP storage) is in progress and will be

reported in the further course of WP3.

5.2.2 Fish

The collected data corresponding to aerobic storage of fish have been presented in detail in the

report of D3.1. The analysis of the data generated during the fish storage experiments under

modified atmosphere packaging (MAP), which were recently completed, is in progress and the

results of this analysis will be reported in the further course of WP3. One recognized (and

solved) issue has been the “bug” in the NIR sensor accompanying software for absorbance

calculation, whereas a second recognized concern has been the light power of the fiber, which

has also been resolved by tripling its power. All adaptations were performed prior to the next

experimental procedures for this specific food commodity (i.e. storage of fish under MAP) as

well as prior to the experimental procedures concerning the additional different sub-use cases.

At this point, we must stress out the importance of the parallel use of additional advanced


Page | 69


spectroscopy methods (i.e. MSI and FTIR) that allowed the assessment of the aforementioned

technical shortcomings.


The results presented in this report refer to the data generated in the course of the rocket

salad storage experiments. The analysis of the data and the interpretation of the results

derived from the experiments on baby spinach and pineapple are in progress, and will be

reported in the further course of WP3.

5.2.3.1 Microbiological data

The results regarding the evolution of microbial populations during storage of ready-to-eat

rocket salad under different temperature conditions are presented in Figure 25. As expected,

the higher the storage temperature the higher the microbial growth rate, and hence, the faster

the spoilage of rocket salad, both from a microbiological (Fig. 25) and a sensorial (data not

shown) perspective. The microbial growth appeared to be similar at the constant temperature

of 8°C and the dynamic (i.e. periodically changing) temperature conditions. Moreover, based on

the microbiological data derived for the specific microbial groups that were examined (data not

shown), it appears that Pseudomonas spp. and molds and yeasts constituted the dominant

spoilage microflora (i.e. specific spoilage organisms) of this fresh produce commodity.

Figure 25: Mean microbial populations during aerobic storage of ready-to-eat rocket salad at different constant (4, 8 and 12°C) and dynamic (8 h at 4°C, 8 h at 8°C and 8 h at 12°C) temperature conditions.


Page | 70


5.2.3.2 Spectral data

The results of the methodology outlined earlier (and described in detail in D3.3), for the all

sensors towards a unified feature selection scheme for spectra data analysis, are presented

herein. Following the RFs-based procedure for feature selection resulted in a feature set that is

very limited compared to the original number of variables (Table 12).

Table 12: Summary of the sensors’ properties, total number of variables and number of

variables resulting after RFs-based feature selection.

Sensor Wavelength/Wavenumber Resolution # variables # selected features

FTIR [1000-3000] cm-1 4 cm-1 2075 94

NIR [1000-1900] nm 1 nm 901 92

VIS [340-850] nm 15 nm 288 79

At this point it should be stated that the aforementioned procedure was applied not only for

the estimation of microbial contamination but also for the time-on-shelf estimation, with the

term “time-on-shelf” referring to the time after the of product’s manufacture (and/or

commercialization). The comparison between the observed (actual) and the estimated

(predicted) microbial contamination and time-on-shelf is illustrated in Figure 26, whereas the

performance metrics of the models correlating each one of these properties on the basis of the

spectral data provided by each one of the tested sensors (FTIR, NIR and VIS) are provided in

Table 13.


Page | 71


Table 13: Linear regression fit properties. Parameters’ values and goodness of fit.

Microbial contamination Slope Offset R2 RMSE

FTIR 0.822 1.215 0.511 0.758

NIR 0.792 1.564 0.945 0.315

VIS 0.623 2.510 0.308 0.649

Time-on-shelf Slope Offset R2 RMSE

FTIR 0.984 14.490 0.958 16.920

NIR 0.725 42.550 0.103 129.400

VIS 0.764 14.710 0.556 43.760


Page | 72


Figure 26: FTIR TVC prediction (A1), FTIR time-on-shelf prediction (A2), NIR TVC prediction (B1), NIR time-on-shelf prediction (B2), VIS TVC prediction (C1), VIS time-on-shelf prediction (C2).

The significance of this decrease in variables, apart from being crucial for training when sample

size is much smaller, it is very significant for several other purposes. The final reduced set of

features is considered de-correlated with minimum redundant information; thus, regression is

more robust than using the full list of features. In addition, the creation/usage of limited sized

featured sets could be very useful in food-specific low cost sensor development since limited

range of wavelengths would be required.


Page | 73


5.3 Feasibility of micro-sensor application for detection of fraud

5.3.1 General food test

In order to make the other partners familiar with the capabilities of the NIR spectrometer, IPMS

had previously carried out some trial measurements to demonstrate the NIR signals of several

different foods measured in transflection geometry. These measurements were done to

illustrate the properties, potential and pitfalls of NIR measurements as well as to demonstrate

typical NIR spectra obtained with the SGS1900 NIR spectrometer. The results of this trial, which

was mainly related to use case 3 but was also instructive to the other use cases, were reported

in detail in D3.1.

5.3.2 Milk powder

Prior to OCC, SMP triplicates were evaluated for consistency per sensor (FLUO, VIS and NIR). In

Figure 27 an example is displayed on a sample outlier (red tetragon) and a faulty triplicate

(yellow tetragons). In order to assess the feasibility of this sub-use case, it is imperative that

divergent measurements are identified and (for the moment) are omitted for OCC. For the in

total 96 spectra per sensor (total amount of spectra: 288), 12 spectra were assigned as outliers

for the FLUO and VIS sensor, and 8 for the NIR sensor. Approximately 8 - 12 % of the

measurements were assigned as outliers. This amount of outliers is too high for a good

application in practice for the PhasmaFOOD sensor. The main reason for this high amount of

outliers is most probable the many degrees of freedom the individual pilot sensors have, i.e. it

is hard for the user of the pilot sensors to acquire replicates of good quality. With the

PhasmaFOOD sensor, this problem should be resolved as the device is optimized for measuring

food stuffs.


Page | 74


Figure 27: Principal component analysis of standard normal variate (SNV) NIR data of 3x32 SMP samples.

Example outlier and faulty triplicate highlighted in red and yellow, respectively.

In Figure 28, the combined results of the food safety related adulterants (nitrogen enhancers)

and the low-value fillers are displayed. The 1152 class distance values resulting from the OCC

approach are transformed (i.e. fused) to a single class distance by using one-class SIMCA. The

displayed line at log class distance value 1 is an arbitrary class line which can be set in the

PhasmaFOOD cloud algorithm in order to tune the desired amounts of false positives and false

negatives. In this case, the line is set at the detection of the lowest amount (0.1 %w/w) of

chemical adulterant detectable (melamine and urea). The SMP outliers (SMP OL) were also

classified in order to confirm or denounce their status as outlier. Indeed, most SMP OLs were

still classified as being abnormal, confirming our observations as stated above in this paragraph.

Interestingly, also a number of SMP OLs were classified as being normal, which confirms the

suitability of OCC for this type of authenticity problems (i.e. even an odd authentic sample or a

poorly conducted measurement is acceptable). For the chemical adulterants, even the lowest

concentration of 0.1 % w/w is separable from the SMP class in the case of melamine and urea.

The other ammonium salts were indisputably classified as adulterated. For the bulk adulterants,

all samples were recognized as adulterated, keeping in mind that no limit of detection can be

deducted from the concentrations used.


Page | 75


Figure 28: One-class SIMCA of the OCC distance values compare to SMP. The line is an arbitrary class separation

line.

Concluding, the combinatorial approach of the PhasmaFOOD FLUO, VIS and NIR sensors by OCC

is promising as all adulterations are detected. Though, still too many outliers were detected and

the dataset was not internally validated as the sample set and amount of replicates per sample

was not high enough. Validation of this result will be performed and reported in D3.7.

5.3.3 Meat

The analysis of data derived from the meat adulteration experiments is in progress and will be

reported in the further course of WP3, in D3.7.


Feasibility will be assessed in the further course of WP3 and will be reported in D3.7.

5.3.5 Edible oils

As mentioned above (in section 3.4.4), IPMS had previously performed a pilot experiment using

the NIR sensor and mixtures of extra virgin olive oil and sunflower oil, and the results of this

experiment were reported in detail in D3.1. Additional feasibility results regarding the use of

the PhasmaFOOD sensors in the detection of edible oils' adulteration will be reported in D3.7.


Page | 76


6 Conclusions and outlook The present D3.2 report, being the second in a series of successive reports, is a compilation of

all additional collected information and experimental data until M18 of the PhasmaFOOD

project, allowing for an updated (compared to that provided in D3.1 report) feasibility

assessment of the proposed use cases (i.e., detection of mycotoxins, spoilage and fraud).

Overall, in the framework of D3.2, elaboration upon planning, sampling and measurement

strategies, standard operating procedures of the VIS and NIR sensors, as well as reference

methods and chemometric protocols has been provided. More specifically, the newly obtained

information presented in this deliverable is pertinent to (i) updated results on mycotoxins'

detection in the sub-use case of maize flour and preliminary results with reference to tree nuts

(almonds) using the VIS sensor; (ii) extensive experimental data reporting along with data

analysis approaches on detection of spoilage of the sub-use cases of fish, meat, fruit (i.e.

pineapple) and vegetables (i.e. rocket salad and baby spinach) using mainly the NIR and VIS

sensors (both visible reflectance and fluorescence measurements); and (iii) experiments

assessing the applicability of the PhasmaFOOD sensors to food fraud detection concerning skim

milk powder and minced meat.

The forthcoming research goals that, in the context of WP3, are anticipated to be completed

and reported in the final deliverable report (D3.7 – M27), involve the continuation and

evolution of data analysis for the already collected data, as well as the initiation and completion

of experiments for the rest of the sub-use cases that have not been investigated yet. With

reference to use case 1 (detection of mycotoxins), tests in the sub-use cases of milk powder and

paprika powder, with the expected measurement set-up being similar to the one used in grain

(i.e. maize) and almond flour, are planned to be conducted. Regarding use case 2 (detection of

spoilage and prediction of shelf-life), next objectives include analysis of data derived from the

fish/minced pork/pineapple/baby spinach experiments (described in the present deliverable

report), as well as conductance of new experiments using the first PhasmaFOOD prototype in

representative sub-use cases. Finally, with regard to use case 3 (detection of food fraud), in

addition to the pilot measurements in edible oils provided in D3.1 report and the already

conducted measurements in skimmed milk powder and minced meat, forthcoming goals

include (i) analysis of collected data on meat adulteration; and (ii) acquisition of additional

experimental data (and corresponding data analysis) on skim milk powder, alcoholic beverages

and edible oils.


Page | 77


References [1] Nemme, K, & Mohammed, A (2017). Mycotoxin occurrence in grains and the role of

postharvest management as a mitigation strategy. A review. Food Control, 78, 412-425.

[2] Zachariasova, M, Dzuman, Z, Veprikova, Z, Hajkova, K, Jiru, M, Vaclavikova, M, Zachariasova,

A, Pospichalova, M, Florian, M, & Hajslova, J (2014). Occurrence of multiple mycotoxins in

European feedingstuffs, assessment of dietary intake by farm animals. Animal Feed Science and

Technology, 193, 124-140.

[3] Borchers, AT, Chang, C, & Gershwin, ME (2017). Mold and human health: a reality check.

Clinical Reviews in Allergy & Immunology, 52, 305-322.

[4] Miedaner, T, Han, S, Kessel, B, Ouzunova, M, Schrag, T, Utz, FH, & Melchinger, AE (2015).

Prediction of deoxynivalenol and zearalenone concentrations in Fusarium graminearum

inoculated backcross populations of maize by symptom rating and near-infrared spectroscopy.

Plant Breeding, 134, 529-534.

[5] Chauhan, R, Singh, J, Sachdev, T, Basu, T, & Malhotra, BD (2016). Recent advances in

mycotoxins detection. Biosensors and Bioelectronics, 81, 532-545.

[6] Hernández, DB, Mishra, RK, Muñoz, R, & Marty, JM (2017). Low cost optical device for

detection of fluorescence from Ochratoxin A using a CMOS sensor. Sensors and Actuators B,

246, 606-614.

[7] Kos, G, Sieger, M, McMullin, D, Zahradnik, C, Sulyok, M, Öner, T, Mizaikoff, B, & Krska, R

(2016). A novel chemometric classification for FTIR spectra of mycotoxin-contaminated maize

and peanuts at regulatory limits. Food Additives & Contaminants: Part, 33, 1596-1607.

[8] Orlov, AV, Burenin, AG, Massarskaya, NG, Betin, AV, Nikitin, MP, & Nikitin, PI (2017). Highly

reproducible and sensitive detection of mycotoxins by label-free biosensors. Sensors and

Actuators B, 246, 1080-1084.

[9] Sharma A, Goud, KY, Hayat, A, Bhand, S, & Marty, JL (2017). Recent advances in

electrochemical-based sensing platforms for aflatoxins detection. Chemosensors, 5, 1;

doi:10.3390/chemosensors5010001.

[10] Toro, PCD, Arévalo, FJ, Fumero, MV, Zon, MA, & Fernández, H. (2016). Very sensitive

electrochemical sensor for moniliformin detection in maize samples. Sensors and Actuators B,

225, 384-390.

[11] Wang, W, Ni, X, Lawrence, KC, Yoon, S-C, Heitschmidt, GW, & Feldner, P (2015). Feasibility

of detecting Aflatoxin B1 in single maize kernels using hyperspectral imaging. Journal of Food

Engineering, 166, 182-192.


Page | 78


[12] Leitzmann, C (1993). Food quality-definition and holistic view. In H. Sommer, B. Petersen, &

P. v. Wittke (Eds.), Safeguarding Food Quality (pp. 3-15). Berlin, Heidelberg: Springer-Verlag.

[13] Hamad, SH (2012). Factors affecting the growth of microorganisms in food. In R. Bhat, A.

Karim Alias, & G. Paliyath (Eds.), Progress in food preservation (pp. 405-427). Oxford: Wiley-

Blackwell.

[14] Koutsoumanis, K, & Nychas, G-JE (2000). Application of a systematic experimental

procedure to develop a microbial model for rapid fish shelf-life predictions. International

Journal of Food Microbiology, 60, 171-184.

[15] Ellis, DI, & Goodacre, R (2001). Rapid and quantitative detection of the microbial spoilage

of muscle foods: current status and future trends. Trends in Food Science and Technology, 12,

414-424.

[16] Karoui, R, & De Baerdemaeker, J (2007). A review of the analytical methods coupled with

chemometric tools for the determination of the quality and identity of dairy products. Food

Chemistry, 102, 621-640.

[17] Dufour, E (2011). Recent advances in the analysis of dairy product quality using methods

based on the interactions of light with matter. International Journal of Dairy Technology, 64,

153-165.

[18] Papadopoulou, OS, Panagou, EZ, Mohareb, F, & Nychas, G.-J. E. (2013). Sensory and

microbiological quality assessment of beef fillets using a portable electronic nose in tandem

with support vector machine analysis. Food Research International, 50, 241–249.

[19] Porep, JU, Kammerer, DR, & Carle, R (2015). On-line application of near infrared (NIR)

spectroscopy in food production. Trends in Food Science and Technology, 46, 211-230.

[20] Xiong, Z, Xie, A, Sun, D-W, Zeng, X-A, & Liu, D (2015). Applications of hyperspectral imaging

in chicken meat safety and quality detection and evaluation: A review. Critical Reviews in Food

Science and Nutrition, 55, 1287-1301.

[21] ICMSF (International Commission on Microbiological Specifications for Foods) (1986).

Sampling plans for fish and shellfish, in: Microorganisms in Foods 2. Sampling for

microbiological analysis: Principles and specific applications, 2nd ed. Blackwell Scientific,

Oxford, pp. 181–196.

[22] Ministére de l’Economie des Finances et du Budget (1988). Marché consummation,

produits vegetaux prêts á l’emploi dits de la “IVemme gamme”: guide de bonnes pratique

hygiéniques. J. Off République Française, 1621, 1-29.

[23] USP Pharmacopeia Convention Food Fraud Database. https://www.foodfraud.org/.


Page | 79


[24] Fernandes, JK, Umebara, T, Lenzi, MK, & Alves, ETD (2013). Image analysis for composition

monitoring. Commercial blends of olive and soybean oil. Acta Scientarum-Technology, 35, 317-

324.

[25] Ottavian, M, Fasolato, L, Serva, L, Facco, P, Massimiliano, B (2014). Food and Bioprocess


[26] Ropodi, AI, Pavlidis, DE, Mohareb, F, Panagou, EZ, & Nychas, G-JE (2015). Multispectral

image analysis approach to detect adulteration of beef and pork in raw meats. Food Research

International, 67, 12–18.

[27] Ropodi, AI, Panagou, EZ, & Nychas, G-JE (2016). Data mining derived from food analyses

using non-invasive/non-destructive analytical techniques; determination of food authenticity,

quality & safety in tandem with computer science disciplines. Trends in Food Science and


[28] Ropodi, AI, Panagou, EZ, & Nychas, G-JE (2017). Multispectral imaging (MSI): A promising

method for the detection of minced beef adulteration with horsemeat. Food Control, 73, 57-63.

[29] Pügner, T., Knobbe, J., & Grüger, H. (2016). Near infrared grating spectrometer for mobile

phone applications. Applied Spectroscopy, 70, 734-745.

[30] Hiperscan GmbH, https://www.hiperscan.com/de/sgs1900

[31] Carstensen, JM, & Hansen, JF (2003). An apparatus and a method of recording an image of

an object. Patent family EP1051660.

[32] Panagou, EZ, Papadopoulou, O, Carstensen, JM, & Nychas, G-JE (2014). Potential of

multispectral imaging technology for rapid and non-destructive determination of the

microbiological quality of beef filets during aerobic storage. International Journal of Food

Microbiology, 174, 1-11.

[33] ISO 12966-4:2015(en) Animal and vegetable fats and oils — Gas chromatography of fatty

acid methyl esters — Part 4: Determination by capillary gas chromatography.

[34] COMMISSION REGULATION (EC) No 273/2008 of 5 March 2008 laying down detailed rules

for the application of Council Regulation (EC) No 1255/1999 as regards methods for the analysis

and quality evaluation of milk and milk products.

[35] Karunathilaka, SR, Farris, S, Mossoba, MM, Moore, JC, & Yakes, BJ (2017). Non-targeted

detection of milk powder adulteration using Raman spectroscopy and chemometrics: melamine

case study. Food additives & contaminants. Part A, Chemistry, analysis, control, exposure & risk

assessment, 34, 170-182.

https://www.hiperscan.com/de/sgs1900


Page | 80


[36] Capuano, E, Boerrigter-Eenling, R, Koot, A, van Ruth, SM (2015). Targeted and untargeted

detection of skim milk powder adulteration by near-infrared spectroscopy. Food Analytical

Methods, 8, 2125-2134.

[37] COMMISSION REGULATION (EC) No 2870/2000 of 19 December 2000 laying down

Community reference methods for the analysis of spirits drinks.

[38] Breiman, L (2001). Random Forests. Machine Learning, 45, 5-32. doi:

10.1023/A:1010933404324.

Documents

DELIVERABLE - phasmafood.eu · Project Full Title: Portable photonic miniaturised smart system for on-the-spot food quality sensing ... (iii) detection of food fraud. In this report,