Cumulative risks from combined exposure to multiple pesticide residues in fruit and vegetables Project Kennis- en modelkoppelingen voor borging voedselveiligheid in de groenten en fruit sector

Dit onderzoek is in opdracht van het Ministerie van LNV uitgevoerd door de Stichting Wageningen Research (WR), business unit Biometris, in het kader van beleidsondersteunend onderzoeksthema Voedselveiligheid (projectnummer TU18039, gunningscode BO-49-002-006). WR is een onderdeel van Wageningen University & Research, samenwerkingsverband tussen Wageningen University en de Stichting Wageningen Research.


Contents
Summary Cumulative intake of mixtures of pesticide residues through consumption of f ruits and vegetables can lead to health risks that are not controlled under the current EU system using only the maximum residue limit (MRL) and acute ref erence dose (ARf D) f or single substances on single f ood products. In a collaboration between the European Commission, the European Food Saf ety Authority and member states, methods have been developed to implement Regulation (EC) No 1107/2009 that states that plant protection products should not have harmf ul health ef f ects, also taking into account possible cumulative and synergistic ef f ects. These methods have been implemented in the Monte Carlo Risk Assessment (MCRA) sof tware which is available f or national and European public institutions to perf orm pesticides mixture risk assessment.
This report describes a web portal that was developed to allow cumulative risk assessment by the Dutch private vegetable and f ruit sector organised in the Foundation Food Compass using MCRA with the monitoring data collected in the Food Compass database.
A case study was perf ormed to combine Food Compass monitoring data f rom the years 2013-2020 with consumption data of children f rom the Dutch national f ood consumption survey. The main interest was to assess the risk of cumulative exposure due to the combined intake of multiple pesticide residues in their diet. If potential risks were observed, it was of interest to know which residues and f oods contributed to such risks and if there were trends over the period of the monitoring. The results of the cumulative assessments were compared to an analysis of limit value exceedances at the level of single measurement results.
The results presented in this report are provisional due to insuf f icient availability of data. Some of the analytical scopes in the Food Compass database could not be linked appropriately to the active substance groups used f or cumulative risk assessment. Food processing steps, such as peeling or juicing of citrus f ruits, are expected to reduce residue levels, but the collection of processing f actors to account f or this in the calculations was incomplete. Limit values are sometimes changed, and recently artif icially low limit values were introduced in the sectoral system to generate alerts f or substances which have been classif ied as genotoxic. Such data will then also have an artif icial impact on the cumulative assessments.
The conventional analysis of single residue measurements showed that 1-3% of residue levels exceeded the MRL throughout the period 2013-2020. However, the f requency of conservatively estimated exceedances of the ARf D using the PRIMo 3.1 model decreased f rom around 3% to below 1% in the same period. It was also f ound that ARf D exceedances did of ten occur without an associated MRL exceedance.
The cumulative assessments indicated that the probability of a critical acute exposure was estimated to be in the range 0.1-0.4% during the period 2013-2019 (the results f or 2020 were af f ected by an artif icial low ARf D value and are theref ore not usef ul to estimate real risk). Nevertheless, the main identif ied risk drivers were occurrence s of chlorpyrif os and chlorpyrif os-methyl in some citrus f ruit products such as juices f or which no account of processing ef f ects was included in the assessment due to lack of validated data. Including validated processing f actors f or citrus f ruit products as they are consumed in practice will result in more realistic critical probabilities which are expected to be lower. It is planned to update the current trend analysis in f urther work.
1 Introduction 1.1 Overview and aim of the study A case study was perf ormed to combine Food Compass monitoring data f rom the years 2013-2020 with consumption data of Dutch children. The main interest is to f ind if there could have been cumulative exposure due to the combined ef f ects of multiple pesticide residues in their diet. If so, it was of interest to know which residues and f oods contributed to such risks and if there were trends over the period of the monitoring. The results presented in this report are provisional due to insuf f icient availability of certain data.
This case study is part of a project to use private sector pesticide residue monitoring data f or cumulative risk assessment using public f ood consumption data and the publicly developed Monte Carlo Risk Assessment (MCRA) web platf orm. For linking private and public data the IPGF portal was developed in the context of a public-private partnership project between WUR Biometris, f inanced f rom public means, and Food Compass, who contributed the monitoring data. IPGF is the abbreviation of 'Impactanalyse Pesticiden in Groenten en Fruit' ('Impact analysis Pesticides in Fruit and Vegetables').
The primary stakeholders of this work are the project partners WUR Biometris and Foundation Food Compass. WUR Biometris aims to build up expertise regarding the linkage of knowledge (seen as data) and models across the internet in an interoperable manner, and to promote the distribution of knowledge about perf orming cumulative risk assessments using the MCRA sof tware. Food Compass aims to be able to perf orm EU compatible cumulative risk assessments on Food Compass samples per period, compare results across periods (trend analysis), to use the results f or risk communication to Food Compass and GroentenFruit Huis colleagues, and possibly to Food Compass participants and/or retail representatives regarding the health impact of cumulative pesticide exposure. A secondary objective is to prepare f or analyses at the sample level by Food Compass participants in the context of quality control or early warning systems.
The longer-term aim is to create better links between private and public knowledge management systems in the interest of open and transparent risk assessment.

Development of a web portal to link private and public data
The IPGF web portal is developed to analyse the concentration levels of substance residues f ound on the f resh f ruit and vegetable samples recorded by Food Compass and assess the human health risk associated with consumption of f ruits and vegetables with such residue concentration levels. Users of the portal are able to evaluate the substance residue concentration levels of F ood Compass samples and 1) compare these levels with legal residue limits (MRLs), 2) compare these lev els with nonstatutory retail requirements, and 3) evaluate the potential human health risk f rom exposure to these concentration levels using dif f erent assessment models. The latter comprises both the single-substance, deterministic IESTI calculations, and more realistic multi-substance, probabilistic cumulative exposure and risk calculations as available in MCRA. In addition, users are able to evaluate the trends of the substance residue concentration levels and their potential associated impact on human health over time.
The IPGF portal retrieves the Food Compass concentration data f rom the Food Compass database, which can be done via the Food Compass web API. These concentration data are validated, curated, and linked with other data as a prerequisite for perf orming the analyses. E.g., linking of f ood codes and curation of analytical scopes. Theref ore, the portal includes a data management module to allow f or data inspection and, if needed, data curation of some identif ied data types.
The potential users of the IPGF portal are Food Compass and GroentenFruit Huis staf f members, Food Compass participants (f or single sample analyses), and interested stakeholders in the Netherlands or in Europe. To allow usage by a broad audience, the portal has been developed in English. Food Compass monitoring data f or the years 2012-2020 were obtained f rom the Food Compass Web API in December 2021. The samples used in the case study were gathered f rom all Food Compass sampling programs: "Monitoring (M)", "EWRS (X)", "NVWA (N)", "Bedrijf seigen monster (E)" and "Aanvullend monster (A)".
Food Compass f ood sample reports were retrieved f or each batch and aligned with the known laboratory scopes. The latter is required because the sample reports contain only inf ormation on the positive substance concentrations and not the substance measurements that were below the detection limit. The Laboratory scope lists provide inf ormation on all substances measured by a given analytical method, including the applicable detection limits. The import step also contains some validation and curation steps to detect and, if possible, restore inconsistencies between the known laboratory scopes and the reported sample substance concentrations. Inconsistencies are reported as critical or non-critical measurement inconsistencies (ref lecting individual substance measurements within a sample) and inconsistencies in linking the reported analysis methods with the known analytical scopes.
For the cumulative exposure assessments, the imported samples are converted to the MCRA concentration data f ormat. The table below shows the results of the sample imports f orming the concentration data f or the batch exposure assessments. In the regulatory system, ARf D values are only set f or active substances that are supposed to have no health ef f ects below a certain threshold. ARf D values are not derived f or active substances with known or presumed health ef f ects due to a nonthreshold mode of action, such as substances that are carcinogenic, mutagenic, or toxic f or reproduction (CMR substances). In practice, this means that such active substances are or will be excluded f rom the market. However, some of these substances were allowed in previous years and are theref ore f ound in historical monitoring data. To obtain a clear alert f or current use, GroentenFruit Huis has decided to include an artif icial very low value f or such substances. Relevant f or the current case study, an artif icial low ARf D of 0.0001 mg/kg bw/day was set f or Chlorpyrif os and Chlorpyrif os-methyl per 13-11-2020. Overview data f or cumulative risk assessment The data required f or cumulative risk assessments in MCRA originate f rom dif f erent sources. The IPGF portal f eeds parts of the data to MCRA, f or example the concentration data, and specif ies the data to be used f or calculation jobs. Some data are already available at MCRA and can be used in assessments by just ref erencing these datasets, f or example the consumption data used f or the assessments. The table below summarizes that data needs f or perf orming the cumulative risk assessments. Health ef f ects and assessment groups data Cumulative exposure assessments were perf ormed f or 15 adverse ef f ects at organ level, with cumulative assessment groups (CAGs) of varying sizes. These CAGs were proposed by Nielsen et al. (2012) in a scientif ic opinion f or the EFSA Pane l on Plant Protection Products and their Residues (PPR). It should be noted, that EFSA has started a process f or more data collection and an updated def inition of CAGs, but this has until now resulted in just three CAGs f or acute ef f ects (EFSA 2019c. For illustrative purposes the 15 CAGs f rom Nielsen et al. (2012) were selected in the context of this case study.
An MCRA ef f ects and assessment groups dataset is created f rom the CAGs proposed by Nielsen et al. (2012). The acute Ef f ects and CAGs at CAG level 1 (organ level) will be used f or the analyses. I.e., an analysis will be done f or each level 1 ef f ect/CAG. The CAG dataset is available on a share in MCRA and a local copy of this data is maintained within the portal f or administration/quality checking. In the current report only the consumption data f or children have been used to have a f irst demonstration of the results. The data has been provided by RIVM to EFSA and has been provided again by EFSA in the f orm of raw primary commodity consumption data (RPC, EFSA 2019a), meaning that the consumptions are expressed in terms of the raw (measured) f ood products.
A number of modelled f oods f ound in the f oods catalogue did not match/align with the RPF consumption data and were theref ore not included in the cumulative assessments (Table 5). The RPC consumption datasets are available f or use on a share on MCRA and a ref erence to these dataset is suf f icient f or using it is an MCRA cumulative exposure analysis f rom the IPGF portal.

2.4.4
Processing f actor data The EFSA database of processing f actors prepared by Scholtz et al. (2018) serves as the basis of the processing f actors used in this case study . However, this dataset contains only a limited amount of substance-f ood combinations. In this project, we noticed in initial assessments that imazalil on citrus f ruits was identif ied as an important risk driver, but that peeling of citrus f ruits is expected to remove most of the imazalil residues. More processing f actors are available in a Dutch database maintained at RIVM in the last updated version of 2020 1 , but the latter database is not organised using the harmonised substance and f ood codes at EFSA and could theref ore not be used automatically. For the analyses reported here, the EFSA processing f actor data were extended with processing f actors f or imazalil in citrus f ood as were available f rom the RIVM database.
For preparing the dataset, the EFSA PARAM codes used by the original processing f actors dataset of Scholtz et al. (2018) were mapped to the Food Compass substance coding system (based on CAS) using the mapping as available in the internal substances catalogue.
A data share in MCRA contains this generated dataset processing f actor dataset, which can be ref erenced f or use in cumulative exposure assessments in MCRA.

2.4.5
Unit variability data The same unit variability are used as used in the Tier II calculations in van Klaveren et al. (2019a) and EFSA (2020a). It should be noted that these studies f ocused on a subset of 30 f ood products and no unit variability f actors are available f or the f ood products not considered by these studies. The unit variability f actors dataset is available f or use on a share on MCRA and a ref erence to this dataset is used within the cumulative exposure analyses in MCRA.

2.4.6
Residue def inition data The cumulative exposure assessments are perf ormed at the level of so -called active substances, which are the substances that are associated with the ef f ects and CAGs and f or which potency inf ormation is assumed to be available. Substance conversions are used f or converting measured substance concentrations (such as sum -substance measurements) to active substance concentrations. The substance conversions are obtained f rom the Food Compass substances hierarchy, which is include d in the substances catalogue.
For each sum-substance that is linked to one or more active substances, substance conversion rules are added to map concentration values of the sum -substances to active substance concentrations. These conversion rules specif y the proportion of measurements of the sum-substance measurements that can be assumed to translate exclusively to a concentration of each active substance, and a conversion f actor to translate the concentration of the sum-substance to a concentration of the active substance. Due to a lack of data, a conversion f actor of 1 is assumed f or all rules and equal proportions of 1/n are assumed f or all active substances linking to the sumsubstance, with n being the total number of substances linking to the sum -substance.
As an example, consider the dithiocarbamates substances in the table below. The sumsubstance (dithiocarbamaten (som als CS2)) links to f our active substance s. For these active substances, f our substance conversion rules are created. Each with a p roportion of 0.25 and a conversion f actor of 1 (see table below).  For each cumulative exposure assessment, a substance conversions dataset is created in this way, uploaded to MCRA, and used in the assessment.

2.4.7
Food translation data (reverse yield f actors) The f ood consumptions of the f ood survey are specif ied at the level of processed raw commodities. Within the cumulative exposure assessments, these consumptions are linked to the measured (raw) f ood products using f ood translations . The f ood translations do not only qualitatively link the processed f oods to the unprocessed/raw f oods, but also include weight correction f actors to translate consumed f ood amounts to equivalent modelled f ood amounts. For the RPC consumption data, the tr anslations data consists of the weight correction f actors due to processing.
The f ood translation dataset is available f or use on a share on MCRA and a ref erence to this dataset is suf f icient f or using it is an MCRA cumulative exposure analysis f rom the IPGF portal. The %ARf D calculation is delegated to the GroentenFruit Huis web service, which computes the %ARf D with the consumption amounts and nominal bodyweights of the critical population and the currently active ARf D value. For the retrospective analyses in the IPGF portal, it is also desirable to compute the %ARf D f or historical samples, using the then-present ARf D value. Theref ore, the %ARf D value received f rom GroentenFruit Huis is recomputed f or historical samples by dividing by the currently active ARf D and multiplying with the ARf D active during the period of sampling.
In the GroentenFruit Huis tool, artif icial low ARf D values were included f or some genotoxic substances such as chlorpyrif os and chlorpyridos-methyl (see section 2.3). Moreover, f or these cases all processing f actors were removed f rom the calculation in the tool.

MRL and ARf D exceedance
In practice, Food Compass is using MRL exceedance as a f ir st screening and ARf D exceedances are registered f or those samples where MRL was exceeded.

Probabilistic cumulative risk assessments
Cumulative ef f ects f rom mixtures of pesticide residues can lead to health risks that are not controlled under the current EU system using only the MRL and ARf D f or single substances. In a collaboration between the European Commission, the European Food Saf ety Authority and member states methods have been developed to assess cumulative exposure and risk (van Klaveren et al. 2019ab; EFSA 2020ab). These methods have been implemented in MCRA (van Klaveren et al. 2019ab).
In the current study we consider risk f or acute health ef f ects as might result f rom consuming f ruit and vegetables f rom the Dutch market against a background of other dietary consumptions of Dutch children. Specif ically, we apply the EC Tier 2 method as was proposed by the European Commission in 2018 and was subsequently adopted by EFSA. For a f ull description of the probabilistic method f or acute health ef f ects see van Klaveren et al. (2019a). Here we only provide a short summary.
Health ef f ects can be grouped according to hierarchical levels. RIVM and EFSA have thus f ar applied cumulative risk assessments f or f our specif ic phenomenological ef f ects, two neurological ef f ects and two ef f ects on the thyroid (level 2; van Klaveren et al. 2019ab, EFSA 2020ab), but have also investigated cumulative risk assessments at the corresponding organ levels, neurological and thyroid (level 1; te Biesebeek et al., 2021). In this case study, cumulative modelling is applied f or 15 cumulative assessment groups (CAGs) def ined at level 1 by Nielsen et al. (2012).

General
The exposure and risk assessment are computed in an acute MCRA risk assessment using the hazard index as the risk metric. Dietary exposures are computed in principle according to EC 2018 Tier 2 settings; meaning that simulated substance residues are generated using a sample-based approach, processing f actors are used in the calculation, and unit-variability is accounted f or in a beta-binomial model using a realistic estimates nature. The active substances of the assessment are obtained f rom data, with a f urther restriction to only the substances f or which an ARf D is available. The hazard characterisations are f ormed by ARf Ds obtained f rom GroentenFruit Huis, which apply to the critical ef f ect that are used as a proxy f or specif ic (organ / CAG level 1) ef f ects in the cumulative exposure assessments. This is similar to the approach f ollowed by te Biesebeek et al. (2021). For each exposure assessment, the most toxic substance (i.e., the substance with the lowest ARf D) was selected as ref erence substance.
Options are available to run the cumulative exposure and risk assessments with or without uncertainty. More specif ically, there are two options f or uncertainty analysis: an uncertainty-test option using only 10 uncertainty analysis cycles (bootstrap cycles), using a reduced population size of 10.000 simulated individuals in the uncertainty cycles, and an uncertainty-f ull option with 100 bootstrap cycles simulating 100.000 individuals in each bootstrap run.

3.2.2
Concentration modelling and occurrence f requencies Concentration modelling is done according to the EC 2018 Tier 2 specif ications. Thus, sample-based concentration modelling is done, non-detects are replaced by 1/2 x LOR and missing values are imputed using occurrence f requency estimates.
Occurrence patterns and f requencies are computed the same way as in the EC 2018 Tier 2 method, except that no substance authorisation data to restrict use percentage up-scaling to authorised uses was used. This is because authorised uses data was not suf f iciently available to use this option.

Extrapolation of f ood samples
No extrapolation was done of f ood samples f or f oods with a limited amount of samples (data poor f oods) f rom other f oods (data rich f oods). The reason f or this is that the extrapolation rules were not suf f iciently available.

Substances conversion
Substances conversion rules are used to translate measured substance concentrations (e.g., of sum substances) to active substance concentrations. The residue def initions are obtained f rom the sum-substance hierarchy inf ormation of the Food Compass substances catalogue. However, substance authorisation inf ormation is not included in the substance conversion, since this inf ormation was not suf f iciently available.
Note that samples can have multiple sample analyses, measured using dif f erent analytical methods. Because of this, it may be possible that there are multiple (conf licting) substance measurements f or the same substance if the substance is by both analytical methods. This may also occur indirectly (via active substance allocation) when one analytical method reports the active substance concentration directly and another analytical method reports a sum-substance concentration that translates that active substance.
As an example, consider the f ollowing example: • A sample can be analysed with both LC-MS and GC-MS.
• Hence, active substance allocation leads to two (possibly conf licting) concentration values f or the active substances bromoxynil (as) and bromoxynil-octanoate on the same sample.
If active substance allocation leads to multiple allocated measurements f or the same substance on the same sample, then the f ollowing procedure is implemented rules f or resolving these inconsistencies: • If all measurements are non-detect, then select the measurement with the smallest LOR.
• If any of the measurements is positive or zero, then take the mean of all positive/zero measurements.
Note that these rules are quite generic and would work quite well also in case there are many measurements f or the same active substance. In practice, one would expect only a f ew (two).

Software: the MCRA platform
The cumulative risk calculations were perf ormed using the Monte Carlo Risk Assessment (MCRA) portal, version 9.1 3 . MCRA 9, also known as the EuroMix Toolbox, is a program f or Monte Carlo Risk Assessment, developed f or RIVM by Wageningen University & Research, Biometris to f acilitate RIVM's tasks f or the Dutch f ood saf ety authority (NVWA) and f or cooperation in international projects (EFSA, EC Research). MCRA 9 was developed in the EuroMix project and in collaborations with EFSA. For acute dietary risk assessment of pesticides, MCRA provides f unctionality to link consumption data f rom a dietary survey and residue occurrence data. Consumption of individual-days are randomly combined with residue levels f or all consumed f oods to produce an estimate of the exposure distribution. Scaled against the ARf D of a substance the exposure distribution can be expressed as a distribution of the hazard quotient (in MCRA termed hazard index, HI), with values above 1 indicating potential risk. For cumulative assessments all residue levels of the substances in an assessment group are scaled by their relative potency f actor (RPF) with respect to a selected index substance. Scaled exposures are summed, and the sum is scaled to HI by dividing by the ARf D of the index substance.
The consumption data and the f iles def ining health ef f ects and cumulative assessment groups are available in MCRA f or the Food Compass user. Many ref inements of the assessment are possible, and some are part of the EC Tier 2 method in this case study. This requires additional data on processing f actors, unit variability f actors and residue def initions. These data are also available in MCRA to the Food Compass user.
The use of MCRA requires a high level of understanding and some degree of experience. To allow the use of MCRA f unctionality f rom other more easily accessible entry points (such as the IPGF platf orm, see next section), an application programming interf ace (API) was created to allow MCRA calculations to be delivered as a web service (WebAPI). For example, external programs can ask MCRA which data is available, send specif ic data to MCRA, ask MCRA to perf orm a specif ic calculation and to send back the results.

Software: the IPGF portal
All analyses of this case study were perf ormed using a beta version of the IPGF web portal (version 2.0.0-beta.1). As mentioned, IPGF portal is a web platf orm specif ically to designed to perf orm human health risk analyses on the Food Compass concentration data. It does so by linking this concentration to other data and delegating model calculations to specif ic modelling services, such as MCRA f or cumulative risk assessment and the GroentenFruit Huis web service f or IESTI and MRL calculations.
The e-inf rastructure of the IPGF portal is depicted in the f igure below. The platf orm can be used to collect data f rom and delegate calculations to dif f erent external web services via Web APIs. These web services are the F ood Compass web service, the GroentenFruit Huis web service, and the MCRA web service. Each service is used f or dif f erent purposes, which is illustrated in the platf orm service inf rastructure diagram of Figure 2. Establishing connections with these services and linking the data f rom multiple sources is theref ore a key aspect of this portal.
The portal presents three types of analyses to analyse the f ood sample s f rom Food Compass: • Single sample analyses: analysis of individual samples, either entered manually or selected f rom all available samples in the FC data (f indable by sample report number). This type of analyses can be perf ormed by all users of the portal.  For the present case study, the batch analysis and the batch comparison analysis are used. Figure 2 shows a screenshot of the batch overview page. On this page, main inf ormation about the batch and the analysis status is shown and f rom this page the user can browse to the various batch result report pages. Figure 3 shows a screenshot of the cumulative exposure assessment results of this batch. For each health ef f ect, it shows the cumulative exposure and conf idence intervals at a specif ied percentile, and the probability of critical exposure (POCE), with its conf idence intervals. As a last example, Figure 4 shows the main results of multiple batches combined in an overview table, which is part of the batch comparison analyses.
In the case study of this report, a batch analysis was done f or each year f rom 2013 to 2020 and the results were extracted f rom the batch analysis reports and the batch comparison report.   Single-substance assessments, business as usual The percentages of samples with exceedance of the MRL or with a calculated IESTI exceeding the ARf D is shown f or the years 2013-2020 in Table 8 and Table 9 and in Figure 5. Large numbers, i.e. 31-41 dif f erent f oods and 45-72 dif f erent substances were involved in the MRL exceedances. Smaller but still quite large numbers, i.e. 5-22 dif f erent f oods and 3-22 dif f erent substances were involved in the ARf D exceedances.
The conventional analysis of single residue measurements showed that 1-3% of residue levels exceeded the MRL throughout the period 2013-2020, without a clear trend.
The most notable change was the gradual decrease of the ARf D exceedance f requency using the PRIMo 3.1 model f rom around 3% in earlier years to 0.7 % in the latest year. Another interesting observation is the low f requency of samples with both MRL and ARf D exceedances (always below 0.5%), meaning that ARf D exceedances may also occur without an accompanying MRL exceedance in the same sample.

4.1.2
Cumulative risk assessments using MCRA The main results f rom the cumulative risk assessments of Dutch children f or the years 2013-2020 is shown in Table 10 and Figure 6. In the table all health ef f ects are listed (if any) which might occur f or at least 0.1% of the population. Two equivalent statistics are shown, the 99.9 th percentile of the %ARf D distribution and the probability of critical exposure (POCE), which is the percentage of persondays with exceedance of the ARf D.
In addition, the f ood-substance combinations that are responsible f or such exceedances are listed. In the f igure each year is represented by the health ef f ect with the highest risk. Due to insuf f icient availability of amongst others processing f actors, these results are provisional.
The cumulative assessments indicated that the probability of a critical acute exposure was estimated to be in the range 0.1-0.4% during the period 2013-2019 (the results f or 2020 were af f ected by an artif icial low ARf D value and are theref ore not usef ul to estimate real risk). Nevertheless, the main identif ied risk drivers were occurrences of chlorpyrif os and chlorpyrif os-methyl in some citrus f ruit products such as juices f or which no account of processing ef f ects was included in the assessment due to lack of validated data.   Only f ew f ood-substance combinations were f ound that contribute to ARf D exceedance when using the probabilistic MCRA model. In order to have a better view on the f oodsubstance combinations that lead to the highest %ARf D values in the simulations, Table 11 and Table 12 list the main contributing f oods, substances and f ood-substance combinations f or the upper 2.5% tail of the cumulative exposure distribution per year f or the health ef f ect with the highest %ARf D values. Figure 7, Figure 8 and Figure 9 show the trends in these main contributions graphically.
The main risk drivers over th ewhole period appeared to be chlorpyrif os in mandarins, apples and oranges. Other combinations that were f ound as occasional risk drivers with more than 10% contribution to the upper exposures were methiocarb in table grapes (in 2013), lambda-cyhalothrin in oranges (in 2019) and chlorpyrif os-methyl in mandarins (in 2020).     In this section detailed results are shown f or the batch analysis of the year 2015, which was the year with the highest observed risk index (excluding 2020, f or which the high risk was due to an artif ical low ARf D). These results and similar results f or other years and health ef f ects are available f or users of the IPGF portal.
As in most years in the trend analysis, developmental ef f ects were identif ied as the health ef f ects of primary concern. The assessment group f or developmental ef f ects consists of 110 active substances, but one of these (f enoxaprop-P-ethyl) excluded f rom the assessment because of a missing ARf D value. For cumulative assessments, oxamyl was selected as the index substance, which means that all exposures are expressed as oxamyl equivalents based on ARf D ratios used as relative potency f actors.
For 54 of the remaining 109 active substances no positive exposure f rom the diet was f ound. Theref ore, the assessment group cumulated the risks f rom ef f ectively 55 substances that were f ound in the diet of the Dutch children. In the simulations, at least one of these substances was present in the diet every day (100% exposure).
The cumulative exposure (in oxamyl equivalents) divided by the ARf D of oxamyl (which happens to be 1 µg/kg/day) specif ies a hazard index (HI) distribution (Figure 10, Figure  11). EC, EFSA, RIVM and NVWA have agreed on using a 99.9% level of protection. Theref ore, the cumulative exposure and HI distributions are evaluated at the 99.9 th percent point (Table 13). Whereas the median estimate of HI is well below 1, the estimate of P99.9 is 1.6 with an uncertainty upper bound of 2.9. This means that f or 0.1% of the children the cumulative exposure is estimated to be 1.6 times the ARf D and could be up to 2.9 times the ARf D. Another way to express these same results is to state that the probability of a critical exposure (POCE) is estimated as 0.33% with an uncertainty upper bound of 0.67%.   .1 -p99.9) in the population. The whiskers indicate a composed confidence interval, the left whisker is the lower 2.5% limit of p0.1, the right whisker is the upper 97.5% limit of p99.9. Zooming in on the individual substance contributions (Figure 12, Figure 13), by f ar the largest contribution to the cumulative risk is seen to come f rom Chlorpyrif os. Zooming in on the individual vegetable and f ruit contributions (Figure 14), the largest contributions to the cumulative risk is seen to come f rom Apples, Mandarins and Oranges. In terms of the consumed products, these are identif ied to be mainly the juiced products, i.e. apple juice, mandarin juice and orange juice. If we consider the most detailed level, we see that indeed Chlorpyrif os in apple, mandarin and orange juices is primarily responsible f or the higher exposures ( Figure  15. Contribution f rom combinations of substance and vegetable and f ruit product as consumed to total (lef t) and upper 0.1% tail (right) cumulative exposure.). A f urther drill-down (Table 14) reveals that chlorpyrif os in orange juice has a processing f actor of 0.025, i.e. the concentrations are assumed to be 40 times lower in orange juice as compared to the raw agricultural product oranges. However, it is also seen that no processing f actors were used f or apple juice, mandarin juice and f or a dif f erent f orm of orange juice (Juicing, Concentration/evaporation). This suggests that exposure f rom these sources is over-estimated. The IPGF portal has been made available f or Food Compass as primary responsible stakeholder f or f ood saf ety of vegetables and f ruits in the Netherlands to perf orm cumulative acute risk assessments. Cumulative assessments, in contrast to IESTI single-substance assessments, address concerns of the EC and the general public about potential mixture ef f ects. Probabilistic assessments are more realistic than simple conservative calculations (IESTI).
It was shown that cumulative assessment provides a usef ul addition f or Food Compass or other private stakeholders to assess the combined risk of multiple chemicals. In this report, batch analyses per year were perf ormed f or investigating the trends in the period 2013-2020. The results presented in this report provide insight in the trends over the years and the high contributing f ood products and substances driving the (mixture) risk.
It is essential to note that the data are not perf ect and there are many aspects where a lack of data (quality) is identif ied, which may lead to a potential bias in the calculations. For a part, this lack of data can be addressed by collecting more data and resolving the quality issues. For complex modelling this seems to be normal, and theref ore we consider optimisation of data organisation as a f orm of a The results of the current analysis highlight chlorpyrif os as the main risk driver. However, Chlorpyrif os (as well as chlorpyrif os-methyl) is now considered to be mutagenic genotoxic and is no longer allowed in the European Union. For these substances an ARf D is no longer available as it is considered mutagenic. This generates a problem f or cumulative assessments which depend on estimation o f the relative potencies. In this study we used ratios of ARf Ds as relative potency f actors. In this report, we theref ore kept using ARf D values, in the f orm as were made available by GroentenFruit Huis. For Chlorpyrif os and Chlorpyrif os-methyl this led to an artif icial high-risk estimate f or the year 2020 because GroentenFruit Huis has set the corresponding ARf Ds to artif icial low values (0.0001 mg/kg bw/d) per 13-11-2020, just to generate alerts. As a consequence, the results of the current trend analysis show an increase in perceived risk in 2020 (note that the ARf D valid at 31-12-2020 was used f or all exposures in the batch). It should be noted that this increase indicates a higher f requency of alerts only and not an increase in real risk (as someone might conclude f rom the 2020 results in chapter 4). For a more realistic comparison between years, it might be usef ul to perf orm calculations with the same ARf D values across the years.
Due to a missing ARf D value, f enoxaprop-P-ethyl was not included in the cumulative calculations. It could be considered to use the ARf D of f enoxaprop-P instead.
In this case study we grouped the pesticides in 15 CAGs corresponding with level 1 (organ level) CAGs as in Nielsen et al. (2012). In this way we covered a variety of potential health ef f ects to illustrate the application of the proposed methodology. However, it should be remarked that these CAGs were not specif ically derived f or acute risk assessment, but rather f or chronic risk assessment, i.e. f ollowing cumulative exposure over time. Specif ic CAGs f or acute health risks have only been derived by EFSA f or two specif ic neurological ef f ects at level 2 (EFSA 2019c, van Klaveren et al. 2019a). The process to def ine more CAGs is ongoing.
Following the approach of te Biesebeek et al. (2021), the hazard characterisations used in the cumulative exposure assessments were ARf Ds applying to the so -called critical ef f ect. For some CAGs, the ARf D of a substance may not relate to the specif ic organ of the CAG, but to an adverse ef f ect on another organ. The use of the ARf D may theref ore be seen as a conservative estimate of an organ-specif ic hazard dose and may lead to an overestimation of the risk.
For ref inement of the calculations perf ormed in this study we identif y the f ollowing types of data that were not yet optimal: 1) Laboratory scopes did not always match with the reported substance measurements, and there is large number of samples and measurement reports with inconsistencies. 2) The set of processing f actors was incomplete, as illustrated f or example in Table  14. This can lead to an overestimation of the exposure, which was also observed in van Klaveren et al. (2019a). 3) In this report, we assumed that our collection of ARf D values represented the valid ARf D values during the period 2013-2020. However, it can be doubted if this collection is complete. The current GroentenFruit Huis web service only provides the most recent ARf D value. It should be discussed how the collection of historical ARf D values can be completed, and in f act, if this is considered a necessary approach. 4) Use of ARf D values to characterise hazard is a conservative approach because in a group of substances the ARf D is related to the the most critical health ef f ect f or each substance individually. If specif ic hazard characterisation data f or each assessment group are available f or all substaces in the group, these could be used (te Biesebeek et al. 2021).
The IPGF portal will be f urther developed in 2022, in line with developments in the methodology of mixture risk assessments by RIVM and EFSA. An update of the trend analysis in this report will be reported. Also, in this study we f ocused on cumulative risk assessments of year batches using the new IPGF portal and MCRA, in comparison to single-sample analyses f or MRL and/or ARf D exceedances. In addition to the analyses described in this report, several other approaches are possible, some of them already implemented and other to be discussed: • Instead of the Monte Carlo simulations, it would also be possible to perf orm probabilistic calculations f or each single sample separately against a background of other samples, and then to summarise the sample results per year. In f act, this was originally the intention, but the computational load turned out to be very high. Further discussion might be needed if this would be a usef ul addition to the current approach.
• Here we have f ocused on trend analysis, i.e. a retrospective assessment. Another potential use of the IPGF portal would be a real-time use, where cumulative risk assessments f or single samples representing product consignments could be used to decide on the acceptability of these consignments.
• In the ongoing discussions, retailers have set additional stringent criteria. In the IPGF portal we have already implemented tabular overviews of the perf ormance of the analysed batches (e.g. per year) against these retail criteria. Based on the results of the cumulative risk assessments, retail requirements could be challenged.