Chronic Dietary Exposure to Pesticide Residues from Apples, Oranges, and Bananas in the HELLANS 2024–2025 Adult Cohort: A Deterministic and Probabilistic Screening Assessment

Abstract

Chronic dietary exposure to pesticide residues links environmental contamination and food safety. This study assessed substance-specific and cumulative exposure from apples, oranges, and bananas among 1318 adults participating in the HELLANS 2024–2025 cohort. Consumption was estimated from a 24-h dietary recall and a food-frequency questionnaire (FFQ), while residue data derived from the Greek national monitoring for 2022–2024. Deterministic chronic exposure was calculated using EFSA PRIMo revision 3.1. Endpoint-specific total margins of exposure (MOETs) were estimated for chronic brain and/or erythrocyte acetylcholinesterase inhibition and hypothyroidism under lower-bound, middle-bound, and upper-bound scenarios. Participant-level cumulative exposure was calculated directly for all 1318 participant records, while uncertainty in monitoring-derived mean concentrations was quantified using 5000 stratified outer-bootstrap replicates. All calculable deterministic estimates were below 100% of the ADI; the highest was phosmet (2.34% ADI). Under the UB scenario, P95 and P99 MOETs were 1324 and 1070 for acetylcholinesterase inhibition and 3944 and 3194 for hypothyroidism; the corresponding P99.9 screening-tail estimates were 928 and 2854. The 95% monitoring-sample uncertainty intervals for the UB P99.9 estimates were 897–954 and 2740–2957, respectively. All point estimates and lower 95% monitoring-sample uncertainty limits exceeded the screening value of 100. Sensitivity analyses using midpoint FFQ frequencies and hypothetical reductions in orange and banana residues did not alter the risk interpretation. Overall, the results did not indicate a chronic screening-level concern from the three assessed fruits within the study cohort, a volunteer sample with high educational attainment, but should not be extrapolated to total dietary exposure or the Greek population.

1. Introduction

Pesticides contribute substantially to crop protection, agricultural productivity, and food security [1]. However, their use may result in residues in food commodities, creating a pathway between environmental contamination and human exposure [2]. For the general population, diet—particularly the consumption of fruits, vegetables, and other plant-derived foods—is a major route of exposure to pesticide residues [3]. Nevertheless, the detection of a residue does not necessarily imply a consumer health risk, which depends on the magnitude of exposure relative to an appropriate toxicological reference value.

Dietary risk assessments have traditionally evaluated active substances individually by comparing estimated exposure with health-based guidance values such as the acceptable daily intake (ADI). Consumers may, however, be exposed simultaneously to multiple residues through different foods. Substance-by-substance assessments may therefore provide an incomplete characterization of combined exposure when several substances affect the same target organ or biological system [4,5,6,7].

Deterministic and probabilistic methods provide complementary information. Deterministic assessments use fixed exposure inputs and are suitable for initial screening, whereas probabilistic assessments characterize distributions of modelled exposure and allow participant variability and sampling uncertainty to be examined separately [3,8,9]. EFSA coordinates pesticide-residue monitoring throughout the European Union, and recent results indicate that most food samples comply with established maximum residue limits and that estimated dietary exposure generally does not raise a concern for consumer health [10]. Dietary exposure assessments can also support the prioritization of active substances and food commodities for monitoring and risk-management actions [11].

To support cumulative pesticide assessment, EFSA has established cumulative assessment groups (CAGs), in which active substances are grouped according to common toxicological effects. EFSA has applied this framework to effects on the nervous system and thyroid gland [12,13] and has continued to prioritize substances and target-organ systems for further evaluation [14,15]. For cumulative risk characterisation, substance-specific exposures can be normalised by dividing each exposure estimate by the endpoint-specific no-observed-adverse-effect level (NOAEL) assigned within the relevant cumulative assessment group. The resulting normalised exposures are summed across substances affecting the same endpoint, and the reciprocal of the cumulative normalised exposure is expressed as the total margin of exposure (MOET) [5,6,7]. European studies have generally reported low chronic dietary exposure relative to health-based guidance values [3,8,9,16]. Greek investigations have similarly documented pesticide residues in commonly consumed foods while generally reporting low estimated dietary risk under the assumptions applied [17,18,19,20,21]. Nevertheless, public concern regarding pesticide residues in Greece emphasizes the need for transparent, context-specific exposure estimates and effective risk communication [22].

Evidence integrating individual-level consumption data, national residue-monitoring results, and both deterministic and probabilistic exposure assessment remains limited in Greece. The present study therefore aimed to characterize chronic dietary exposure to pesticide residues from apples, oranges, and bananas among adults participating in the HELLANS 2024–2025 cohort. Cohort-specific consumption and body-weight information was combined with Greek residue-monitoring data from 2022–2024. Substance-specific exposure was assessed using EFSA PRIMo revision 3.1, while cumulative risk characterisation focused on the established EFSA chronic endpoints of brain and/or erythrocyte acetylcholinesterase inhibition and hypothyroidism. The study also examined the influence of left-censored monitoring results and uncertainty in monitoring-derived mean concentrations.

2. Materials and Methods

2.1. Study Design and Consumption Assessment

Data were collected through an online questionnaire administered via the European Commission’s EUSurvey platform between May 2024 and April 2025 as part of the HELLANS 2024–2025 study. Adult volunteers were recruited by convenience sampling through email invitations, social media, websites, and online news platforms, yielding 1318 completed responses. Participation was voluntary and anonymous, and informed consent was obtained from all participants. The study protocol was approved by the Research Ethics and Deontology Committee of the Hellenic Mediterranean University (approval No. 12656, 16 May 2024) and complied with the Declaration of Helsinki and the General Data Protection Regulation.

The questionnaire was developed based on previous studies [23,24] and comprised sections on sociodemographic characteristics, attitudes towards pesticide residues and food safety, and food consumption. The variables used in the present exposure assessment were participant body weight and the consumption of apples, oranges, and bananas. These fruits were selected because they were among the most frequently consumed fruits in the cohort, were broadly available throughout the year, and had sufficient residue-monitoring data.

Food consumption was assessed using a single 24 h dietary recall (24HR) and a short food-frequency questionnaire (FFQ). The 24HR recorded the foods and corresponding amounts consumed during the preceding day. It was completed online without interviewer assistance, and NHANES 2019–March 2020/FNDDS data were used to select reference portion weights [25]. The research team prepared, weighed, and photographed the corresponding fruit portions, and participants selected the image closest to the amount consumed. The image library was standardized by direct weighing but was not formally validated.

The FFQ recorded consumption frequency during the preceding year as never, less than one day per month, 1–3 days per month, one day per week, 2–3 days per week, 4–5 days per week, or almost daily/daily. The upper bound of each interval was used as a conservative estimate, corresponding to 0, 12, 36, 52, 156, 260, and 365 days/year, respectively. For participant (j) and fruit (f), the corresponding probability of consumption was calculated as follows:

$$P_{\mathit{fj}}={\displaystyle \frac{D_{\mathit{fj}}}{365}}$$

where D_fj is the assigned annual number of consumption days.

Because the FFQ did not collect portion sizes, expected long-term daily intake was estimated as follows:

$$I{R}_{\mathit{fj}}=P_{\mathit{fj}}\times \overline{A_f}$$

where P_fj is the consumption probability for participant (j) and fruit (f), and $\overline{A_f}$ is the mean amount consumed among 24HR consumers of that fruit. The complete participant row—body weight and expected intakes of all three fruits—was retained as a joint observation in all participant-level analyses. These values represent modelled expected daily intake rather than individual usual-intake distributions because only one 24HR was available and between-person variation in portion size was not fully represented [26,27,28].

Data collection covered a complete annual cycle. However, because the volunteer sample was not nationally representative and population weights were not applied, all consumption and exposure distributions were interpreted as specific to the HELLANS cohort.

2.2. Pesticide Residue Data and Treatment of Left-Censored Results

Pesticide residue data for apples, oranges, and bananas were obtained from the Greek national monitoring program for 2022–2024 [29,30,31]. Sampling strategy was identified from the EFSA Standard Sample Description metadata. Following EFSA cumulative-risk assessment practice [5], only samples coded as objective (ST10A) or selective (ST20A) sampling and belonging to eligible official programme types were retained. Suspect samples (ST30A) and samples from EU increased-control or national third-country import-control programmes were excluded. Of 1072 primary samples, 997 were retained (539 apple, 329 orange, and 129 banana samples); the 75 excluded samples and the monitoring-data audit are documented in Supplementary Data S1. The annual datasets were harmonised with respect to commodity, active-substance identity, result value, quantification status, and the reported limit of quantification (LOQ). Substance names and official residue-definition codes referring to the same active substance were consolidated before calculating occurrence and concentration statistics. Annual datasets were pooled separately by commodity; thus, each year contributed in proportion to its available number of samples. Substance–commodity combinations are reported in Table A2 only when at least one result was quantified.

For the deterministic PRIMo assessment, results below the LOQ were assigned LOQ/2. For the cumulative MOET assessment, three separate scenarios were evaluated:

• Lower bound (LB): results below the LOQ were assigned zero;
• Middle bound (MB): results below the LOQ were assigned LOQ/2;
• Upper bound (UB): results below the LOQ were assigned the LOQ.

Arithmetic mean concentrations were calculated separately for every substance–commodity–scenario combination.

No peeling or processing factors were applied because reliable, directly matching factors were not consistently available for all substance–commodity combinations. The assessment was therefore conducted as a whole-commodity screening analysis. Comparable analyses have noted that the application of suitable processing factors would generally be expected to reduce estimated exposure [32].

2.3. Deterministic Exposure Assessment

Chronic deterministic exposure was assessed using the official EFSA Pesticide Residue Intake Model (PRIMo), revision 3.1 [33]. The unrounded deterministic calculation audit is provided in Supplementary Data S1, and a completed three-fruit PRIMo worked example is provided in Supplementary File S1. Study-specific arithmetic mean consumption and body-weight values and arithmetic mean residue concentrations from the pooled 2022–2024 monitoring data were entered into the relevant input cells. Exposure was aggregated across apples, oranges, and bananas for each active substance and expressed as a percentage of the substance-specific ADI. Current EU ADIs were used when available [34]. When no current ADI could be identified, the occurrence result was retained, as shown in Table A2, but the deterministic %ADI is reported as “not available” and was not calculated.

2.4. Endpoint-Specific Chronic Cumulative Exposure and Risk Characterisation

Cumulative risk characterisation was based on established EFSA cumulative assessment groups (CAGs) and endpoint-specific no-observed-adverse-effect levels (NOAELs). Two chronic endpoints were evaluated: (i) brain and/or erythrocyte acetylcholinesterase inhibition and (ii) hypothyroidism. Active-substance assignments and endpoint-specific reference points were obtained from the corresponding EFSA nervous-system and thyroid CAG reports [12,13] and are documented in Table A4. Assignments were non-exclusive, and only substances quantified in at least one assessed commodity and having both an eligible CAG assignment and an endpoint-specific NOAEL were included.

The primary estimand was the MOET corresponding to specified percentiles of the empirical participant-level cumulative normalised-exposure distribution. Within each censoring scenario, commodity–substance arithmetic mean concentrations were fixed, and exposure was calculated for all 1318 participant records, thereby preserving the observed joint distribution of body weight and consumption of the three fruits. Sex was not used for stratification, and sex-specific exposure distributions were not estimated.

Cumulative normalised exposure (NE) was calculated as follows:

$$\mathrm{NE}=\mathsf{\Sigma }\left[\right(\overline{\mathrm{C}}\times \mathrm{I}/1000)/(\mathrm{BW}\times \mathrm{NOAEL}\left)\right],$$

where $\overline{\mathrm{C}}$ is the scenario-specific arithmetic mean concentration (mg/kg), I is the participant’s expected daily intake of a fruit (g/day), BW is body weight (kg), and NOAEL is the endpoint-specific reference point (mg/kg bw/day). For each endpoint and scenario, the P50, P95, P99, and P99.9 of the participant-level normalised-exposure distribution were calculated from all 1318 values. The total margin of exposure (MOET) at exposure percentile P was then calculated as follows: MOET = 1/Qp (NE), where Qp is the corresponding percentile of cumulative normalised exposure across the 1318 participant records.

Thus, the reported MOET at a given exposure percentile is not the same percentile of the participant-specific MOET distribution [35]. EFSA cumulative risk assessments report MOETs at defined exposure percentiles and use a value of 100 at P99.9 as the threshold for regulatory consideration [5,35]. LB, MB, and UB were analysed separately to assess sensitivity to the treatment of left-censored monitoring results [5,35,36].

2.5. Outer-Bootstrap Analysis of Monitoring-Sample Uncertainty

Sampling uncertainty in the commodity–substance arithmetic mean concentrations was quantified using a 5000-replicate outer non-parametric bootstrap implemented in Microsoft Excel 365 with XLRisk version 1.0 [37]. Separate formula-based workbooks were used for the LB, MB, and UB scenarios (Supplementary File S1), with a fixed random seed of 20260826. XLRisk was configured for ordinary Monte Carlo sampling with Latin hypercube sampling disabled. Within each active substance × commodity × year stratum, analysed monitoring observations were resampled with replacement while preserving the original stratum sample size; the resulting resamples were pooled across years, thereby retaining the original sample-size weighting of the 2022–2024 monitoring data.

In each replicate, scenario-specific commodity–substance arithmetic mean concentrations were recalculated and propagated through the cumulative exposure model using all 1318 participant records. Participant inputs, endpoint assignments, and endpoint-specific NOAELs were held fixed. The P50, P95, P99, and P99.9 of the participant-level normalised-exposure distribution and the corresponding MOETs were recalculated for each replicate.

Iteration-level MOET outputs were summarised in JASP version 0.98.1 [38], with quantile type 7 selected for percentile estimation. For each endpoint, scenario, and exposure percentile, the 95% monitoring-sample uncertainty interval (UI) was defined as the 2.5th–97.5th percentile interval of the bootstrap MOET distribution [7,35]. These intervals quantify sampling uncertainty in the monitoring-derived mean concentrations, conditional on the observed participant-level body weight and consumption inputs and the specified toxicological model. They do not incorporate uncertainty in the participant sample, consumption model, residue definitions, CAG membership, endpoint selection, or toxicological reference points and should not be interpreted as intervals for the overall uncertainty of the exposure assessment.

Bootstrap convergence was assessed by comparing the 2.5th and 97.5th percentile estimates after 500, 1000, 2500, and 5000 replicates (Supplementary Data S2).

Sparse monitoring strata are documented in Supplementary Data S1. Iteration-level MOET outputs, JASP percentile summaries, and convergence diagnostics are provided in Supplementary Data S2.

2.6. Sensitivity Analyses

Two sensitivity analyses were performed to examine model assumptions. First, the FFQ frequency categories were recoded using category midpoints rather than the upper-bound values used in the primary analysis. The sensitivity coding was 0, 6, 24, 52, 130, 234, and 338 days/year for never, <1 day/month, 1–3 days/month, 1 day/week, 2–3 days/week, 4–5 days/week, and almost daily/daily, respectively. Fruit-specific mean 24HR portion amounts and individual body weights were held fixed. Participant expected daily intakes were recalculated and propagated through the deterministic %ADI and endpoint-specific MOET calculations, while residue concentrations, endpoint assignments, and NOAELs remained unchanged.

Second, to quantify the potential influence of using whole-commodity residue concentrations rather than edible-portion concentrations for oranges and bananas, the UB cumulative assessment was repeated after applying hypothetical 25%, 50%, and 75% reductions to the orange and banana mean residue concentrations, with apple concentrations unchanged. These scenarios were intended solely as an illustrative quantitative sensitivity analysis of possible edible-portion effects and should not be interpreted as validated commodity- or substance-specific processing factors. The sensitivity results were compared with the primary point estimates; the monitoring-sample outer bootstrap was not repeated because the objective was to examine the effect of fixed model-input assumptions. Complete results are provided in Supplementary Data S2.

3. Results

3.1. Study Sample and Participant Characteristics

The cohort comprised 1318 adults, including 763 females (57.9%) and 555 males (42.1%). The largest age groups were 45–54 years (40.9%) and 55–64 years (25.6%), while 84.4% of participants had completed tertiary education. Participants were distributed across Northern (35.8%), Central (31.9%), and Southern Greece (32.2%); complete sociodemographic characteristics are presented in Table A1. Mean body weight was 78.8 kg (95% bootstrap CI: 77.9–79.7 kg). Participant body weight and the modelled expected intakes of all three fruits were retained jointly in the endpoint-specific cumulative calculations and were held fixed during the monitoring-data bootstrap.

3.2. Selection and Consumption of the Assessed Fruits

Apples, oranges, and bananas were selected because of their relatively high consumption frequencies, broad year-round availability, and sufficient residue-monitoring samples. Modelled expected daily consumption relative to body weight was highest for oranges, followed by apples and bananas (Table 1). These values represent FFQ-frequency-weighted expected daily intakes rather than empirical distributions of individual usual intake. Summary statistics for body weight and the fruit-consumption inputs are provided in Table S2 within Supplementary Data S1.

Table 1. Modelled expected daily consumption of the three assessed fruits in the study cohort (N = 1318).

Fruit	Mean, g/kg bw/day (95% Bootstrap CI)	P97.5, g/kg bw/day
Apples	1.119 (1.063–1.178)	3.614
Oranges	1.437 (1.369–1.506)	4.411
Bananas	1.027 (0.979–1.077)	3.165

Note: Consumption estimates combine participant-specific FFQ probabilities with the mean amount reported by consumers in the 24 h recall. Accordingly, P97.5 refers to the distribution of modelled expected daily intake within the study cohort. CI, 95% bootstrap confidence interval based on 10,000 resamples (JASP version 0.98.1).

3.3. Occurrence of Pesticide Residues

Analytical coverage varied by substance and commodity: analyte-specific denominators ranged from 2 to 539 samples for apples, 15 to 329 for oranges, and 17 to 129 for bananas. Consequently, quantification percentages were interpreted together with their corresponding analysed-sample counts. Sparse analytical coverage was observed for propamocarb, for which only two samples were analysed and both were quantified, and for fosetyl-Al and dithianon, for which 12 samples each were analysed and 5 and 4 results, respectively, were quantified. Among substances with broad apple coverage, acetamiprid was quantified in 133/539 samples (24.7%), boscalid in 106/474 (22.4%), etofenprox in 97/539 (18.0%), and tebuconazole in 96/539 (17.8%). The highest apple concentrations were 4.915 mg/kg for imazalil and 4.390 mg/kg for pyrimethanil.

In oranges, pyriproxyfen and imazalil were the most frequently quantified substances with broad analytical coverage, occurring in 51/329 (15.5%) and 46/329 (14.0%) samples, respectively. Fosetyl-Al was quantified in 2/15 samples (13.3%), whereas phosmet was quantified in 19/326 (5.8%). The highest orange concentrations were 2.700 mg/kg for thiabendazole and 1.880 mg/kg for imazalil.

In bananas, copper was quantified in 16/17 analysed samples (94.1%), although its analytical coverage was limited. Azoxystrobin, thiabendazole, and pyriproxyfen were quantified in 94/129 (72.9%), 80/129 (62.0%), and 35/129 (27.1%) samples, respectively. Maximum concentrations were 2.300 mg/kg for copper, 1.174 mg/kg for thiabendazole, and 0.790 mg/kg for azoxystrobin. Complete analyte-specific occurrence data and LB, MB, and UB arithmetic mean concentrations are presented in Table A2.

3.4. Substance-Specific Deterministic Chronic Exposure Assessment

All substance-specific deterministic estimates were below 100% of the corresponding ADI (Table A3). Phosmet had the highest aggregate exposure across the three fruits at 2.343% ADI, followed by lambda-cyhalothrin (0.800%), acetamiprid (0.464%), imazalil (0.363%), tau-fluvalinate (0.332%), pyrimethanil (0.254%), and fenpropimorph (0.210%). The highest fruit-specific values were 1.024% ADI for phosmet from apples, 1.319% ADI for phosmet from oranges, and 0.210% ADI for fenpropimorph from bananas.

The %ADI values were interpreted separately by active substance and were not summed across toxicologically unrelated substances. Chlorpyrifos, piperonyl butoxide, dithiocarbamates measured as CS₂, and total measured copper were retained in the occurrence assessment but were not expressed as %ADI: suitable current ADIs were unavailable for chlorpyrifos and piperonyl butoxide, the non-specific CS₂ result could not be assigned to an individual dithiocarbamate active substance, and pesticide-source attribution was not possible for total copper. However, chlorpyrifos was retained in the AChE MOET analysis because an endpoint-specific NOAEL was available for this effect. Complete fruit-specific and three-fruit PRIMo estimates are presented in Table A3.

3.5. Endpoint-Specific Chronic Cumulative Exposure and Monitoring-Sample Uncertainty

Endpoint-specific MOETs decreased from the LB to the MB and UB scenarios, consistent with increasing substituted concentrations for left-censored results (Table 2). Greater interpretive emphasis was placed on P95 and P99 because these percentiles are less dependent on the most extreme participant-level exposure observations than P99.9. Under the UB scenario, the AChE MOETs were 1324 at P95 and 1070 at P99, while the corresponding hypothyroidism MOETs were 3944 and 3194. The UB P99.9 screening-tail MOETs were 928 for AChE inhibition and 2854 for hypothyroidism. Complete percentile-specific results for all scenarios are provided in Table 2 and Table A5. Among the 67 active substances/residue definitions quantified in at least one assessed commodity, 22 met the criteria for inclusion in at least one of the two endpoint-specific cumulative assessments; the remaining 45 lacked an eligible EFSA CAG assignment and/or endpoint-specific reference point for these endpoints.

Table 2. Endpoint-specific chronic cumulative total margins of exposure (MOETs) under the lower-bound, middle-bound, and upper-bound scenarios, with monitoring-sample outer-bootstrap uncertainty at the P99.9 exposure level.

Endpoint-Specific Cumulative Assessment Group	Scenario	MOET at P50 Exposure	MOET at P95 Exposure	MOET at P99 Exposure	MOET at P99.9 Exposure	95% Monitoring-Sample UI for MOET at P99.9
Chronic brain and/or erythrocyte acetylcholinesterase inhibition	LB	55,270	18,121	15,021	13,023	8317–23,276
Chronic brain and/or erythrocyte acetylcholinesterase inhibition	MB	7395	2482	1995	1732	1617–1832
Chronic brain and/or erythrocyte acetylcholinesterase inhibition	UB	3964	1324	1070	928	897–954
Hypothyroidism	LB	48,656	17,434	13,818	12,065	10,216–14,244
Hypothyroidism	MB	19,649	6482	5200	4666	4354–4971
Hypothyroidism	UB	12,251	3944	3194	2854	2740–2957

Notes: MOET, total margin of exposure; UI, uncertainty interval. LB, MB, and UB assign results below the LOQ to zero, LOQ/2, and the LOQ, respectively. MOETs are the reciprocals of the corresponding percentiles of cumulative normalised exposure. The 95% UIs were derived from 5000 stratified outer-bootstrap replicates of the monitoring observations and are conditional on the observed participant-level consumption and body-weight inputs and the specified toxicological model; they do not represent overall uncertainty in the exposure assessment. Complete percentile-specific UIs are provided in Table A5. A MOET of 100 at P99.9 was used as the screening reference value.

For the UB scenario, the 95% monitoring-sample UIs were 1286–1369 and 1031–1101 for the AChE P95 and P99 MOETs, respectively, and 3797–4092 and 3061–3317 for the corresponding hypothyroidism estimates. The P99.9 UIs were 897–954 for AChE inhibition and 2740–2957 for hypothyroidism. All point estimates and lower monitoring-sample UI limits remained above 100. All 5000 bootstrap replicates yielded valid MOET estimates for every reported combination, and the maximum relative change in either UI bound between 2500 and 5000 replicates was <1% across all censoring scenarios (Supplementary Data S2).

3.6. Sensitivity Analyses Results

Recoding the FFQ categories to midpoint frequencies reduced the mean modelled daily intake of apples, oranges, and bananas by 10.9–11.8% relative to the primary upper-bound frequency coding. The highest deterministic estimate, phosmet, decreased from 2.343% ADI to 2.069% ADI. Endpoint-specific MOETs increased under the midpoint coding. In the UB scenario, the AChE MOETs at P95, P99, and P99.9 increased from 1324, 1070, and 928 to 1467, 1155, and 1002, respectively; the corresponding hypothyroidism MOETs increased from 3944, 3194, and 2854 to 4312, 3451, and 3088. The interpretation therefore remained unchanged (Supplementary Data S2).

The hypothetical edible-portion residue-reduction analysis also moved the MOETs away from the screening value of 100. Under UB, a 50% reduction in orange and banana mean residue concentrations increased the AChE MOETs at P95, P99, and P99.9 to 1848, 1481, and 1284, respectively, and the corresponding hypothyroidism MOETs to 4614, 3789, and 3322. The 25% and 75% reduction scenarios showed the same direction of change. These scenarios are illustrative and should not be interpreted as validated processing-factor corrections (Supplementary Data S2).

4. Discussion

This study combined recent Greek residue-monitoring data with cohort-specific consumption and body-weight information to characterize chronic dietary exposure to pesticide residues from apples, oranges, and bananas. The assessment produced two complementary findings. First, all substance-specific deterministic exposures estimated with EFSA PRIMo revision 3.1 were well below 100% of the corresponding ADI. Second, the endpoint-specific cumulative MOET estimates at P95 and P99, as well as the P99.9 screening-tail estimates and their lower 95% monitoring-sample uncertainty limits, remained above the screening reference value of 100 under the LB, MB, and UB scenarios. These findings do not indicate a chronic screening-level concern from the three assessed fruits within this cohort under the assumptions applied, but they should not be interpreted as evidence of zero risk or as an assessment of total dietary exposure.

The substance-specific deterministic PRIMo assessment and the endpoint-specific cumulative assessment served different purposes. EFSA PRIMo revision 3.1 estimated substance-specific chronic exposure relative to each substance’s ADI [33], whereas the cumulative analysis normalised each substance–commodity exposure by the endpoint-specific NOAEL assigned within an established EFSA chronic cumulative assessment group [12,13]. Direct numerical agreement between %ADI and MOET was therefore neither expected nor required. All calculable deterministic estimates remained well below 100% ADI. Chlorpyrifos and piperonyl butoxide were not expressed as %ADI because suitable current ADIs were unavailable, while dithiocarbamates measured as CS₂ could not be attributed to an individual active substance, and total measured copper could not be attributed specifically to pesticide use. These distinctions illustrate why regulatory status, residue occurrence, residue definition, source attribution, and toxicological eligibility must be documented separately.

The cumulative results were sensitive to the treatment of left-censored measurements, as expected. MOETs decreased progressively from LB to MB and UB because substituting higher values for results below the LOQ increased normalised exposure. The reduction was particularly marked for the AChE endpoint, showing that substituted values materially influenced the estimated distribution even though the risk interpretation did not change. Reporting all three scenarios is therefore preferable to relying on LOQ/2 alone because it makes the influence of left-censoring assumptions explicit.

The outer bootstrap addressed sampling uncertainty in the monitoring means, which is distinct from participant-level exposure variability. In each of 5000 replicates, monitoring observations were resampled within active substance × commodity × year strata, scenario-specific arithmetic mean concentrations were recalculated, and the endpoint-specific exposure model was rerun while participant records, CAG assignments, and NOAELs were held fixed. The resulting intervals are therefore conditional on the observed participant-level body weight and consumption inputs and the specified toxicological model. They do not quantify uncertainty in the participant sample, consumption model, residue definitions, end-point assignments, NOAELs, or other structural assumptions and should not be interpreted as overall uncertainty intervals for the exposure assessment.

Commodity-specific patterns were also evident in the deterministic assessment. Phosmet provided the highest fruit-specific %ADI estimate for apples (1.024%) and oranges (1.319%), whereas fenpropimorph provided the highest estimate for bananas (0.210%). These rankings cannot be inferred from quantification frequency alone because exposure also depends on the mean concentration, fruit consumption, body weight, and toxicological reference value. Detection or quantification frequency should therefore not be interpreted as a direct measure of health risk.

No peeling or processing factors were applied in the primary analysis; residue concentrations therefore referred to the monitored commodities rather than exclusively to their edible portions. This assumption is particularly relevant to oranges and bananas and may overestimate edible-portion exposure when residues are concentrated in the peel. Published processing studies indicate that washing, peeling, and juice production can substantially reduce pesticide residues in citrus, although the magnitude varies among active substances and processing conditions [32,39]. Because directly applicable sub-stance–commodity-specific factors were unavailable for all relevant combinations, generic processing corrections were not introduced into the primary model. Instead, the sensitivity analysis applied hypothetical 25%, 50%, and 75% reductions to orange and banana residues under UB. All three scenarios increased the endpoint-specific MOETs; for example, a 50% reduction increased the AChE P95/P99/P99.9 MOETs from 1324/1070/928 to 1848/1481/1284 and the corresponding hypothyroidism MOETs from 3944/3194/2854 to 4614/3789/3322. These results illustrate the likely direction and magnitude of edible-portion effects but are not validated processing-factor estimates.

Previous European studies using Belgian, Danish, Hungarian, and Greek data generally reported low chronic dietary exposure under their respective study conditions [3,7,8,9,10,16,17,18,19,20,21]. These findings provide qualitative context only. Direct numerical comparisons are not appropriate because the studies differed in their populations, foods covered, monitoring periods, processing assumptions, treatment of left-censored results, consumption inputs, model structures, toxicological groupings, and risk metrics. In particular, earlier HI estimates based on sums of ADI-normalised HQs are not directly comparable with the endpoint-specific MOETs reported here. Although probabilistic methods have also been used in previous cumulative assessments [5,6,7,8,9,16], differences in model design preclude direct numerical comparison. The present findings should therefore be interpreted within the HELLANS cohort and the defined three-fruit exposure scenario and not as confirmation of or a benchmark against previous estimates.

The joint empirical participant records preserved the observed relationship between body weight and consumption of the three fruits. Cross-commodity dependence in residue concentrations could not be evaluated because monitoring samples were not linked across fruits by consumer, purchase occasion, origin, or batch. Accordingly, the reported uncertainty intervals quantify monitoring-sample uncertainty conditional on the observed participant data and specified exposure model rather than uncertainty from a full two-dimensional probabilistic analysis.

Interpretation of occurrence data also depends on analyte-specific analytical coverage. The pooled monitoring files did not contain an identical number of valid results for every substance, commodity, and year; consequently, occurrence percentages and mean concentrations were calculated only among samples in which the substance was analysed. Analytical coverage varied markedly among substance–commodity combinations, and quantification percentages should therefore be interpreted together with the corresponding numbers analysed. For example, propamocarb was quantified in both two apple samples in which it was analysed, whereas cyantraniliprole was quantified only once despite much broader analytical coverage. These contrasting cases illustrate that quantification frequency alone does not provide a reliable indication of residue prevalence or exposure relevance.

Several limitations should also be considered. The HELLANS cohort was a volunteer sample without population weighting with high educational attainment, which may have systematically influenced fruit-consumption patterns and, consequently, the estimated exposure distribution; the estimated intake and MOET distributions should not be generalized to all Greek adults. Probabilities derived from FFQ were combined with the mean amount consumed from a single 24HR, producing modelled expected long-term intake rather than complete individual usual-intake distributions [26,27,28]. The self-administered 24HR used a study-specific library of directly weighed fruit portions but did not implement the complete NHANES Automated Multiple-Pass Method, interviewer assistance, or formal independent validation of the image library [25]. Using the upper bound of each FFQ frequency interval may overestimate modelled long-term intake relative to midpoint coding; however, the midpoint sensitivity analysis did not alter the overall risk interpretation. Caution is warranted when interpreting P99.9. Given the cohort size (n = 1318), this extreme percentile is strongly influenced by a small number of high-exposure participant records. P99.9 was retained because of its relevance to the EFSA cumulative-risk screening framework, but greater interpretive emphasis was placed on P95 and P99 because they are less sensitive to the most extreme observations in the empirical exposure distribution.

The assessment was restricted to three fruits and therefore does not characterize total dietary exposure; exclusion of other foods may underestimate overall exposure despite conservative assumptions for the included commodities. Monitoring data from 2022–2024 were pooled in proportion to the available sample numbers [29,30,31], preserving the empirical dataset but not modelling temporal trends or uncertainty in annual sampling intensity. Substances not quantified in any assessed commodity and substances without an eligible endpoint assignment or reference point could not contribute to the corresponding MOET. Accordingly, the UB scenario is an upper-bound substitution for left-censored observations within the included endpoint model, not an upper bound for every pesticide that may occur in the total diet.

Despite these limitations, the assessment integrated national monitoring data with participant-level dietary and body-weight information, distinguished substance-specific deterministic exposure from endpoint-specific cumulative risk characterization, and explicitly addressed uncertainty and key modelling assumptions through outer-bootstrap and sensitivity analyses. Inclusion decisions were based on established EFSA endpoint assignments and reference points rather than contribution thresholds. Overall, these features provide a transparent screening characterization of chronic cumulative exposure from apples, oranges, and bananas in the HELLANS cohort while clearly defining the boundaries of inference and the remaining sources of uncertainty.

5. Conclusions

Pesticide residues were quantified in apples, oranges, and bananas, but their presence alone does not imply a consumer health risk. All substance-specific deterministic chronic exposure estimates calculated with EFSA PRIMo revision 3.1 were substantially below 100% of the corresponding ADI. In the endpoint-specific cumulative assessment, the P95 and P99 MOET estimates, as well as the P99.9 screening-tail estimates and their lower 95% monitoring-sample uncertainty limits, remained above the screening reference value of 100 under the LB, MB, and UB scenarios. The most conservative primary result was the UB P99.9 MOET for acetylcholinesterase inhibition, at 928 (95% monitoring-sample UI: 897–954).

Overall, the findings did not indicate a chronic screening-level dietary concern from pesticide residues in the three assessed fruits within the HELLANS cohort, under the assumptions applied. The interpretation was unchanged in the FFQ midpoint and hypothetical orange/banana residue-reduction sensitivity analyses. Because the cohort was a non-representative volunteer sample with high educational attainment, the findings should not be extrapolated to the broader Greek adult population or to total dietary exposure. Future assessments should extend food coverage, nationally representative consumption data, more uniform analytical coverage, and more fully characterize uncertainty in participant-derived inputs and toxicological assumptions.