D2.5 – Open source model for probabilistic human health risk assessment

Project ID N°: 101036449 Call: H2020-LC-GD-2020-3 Topic: LC-GD-8-1-2020 - Innovative, systemic zero-pollution solutions to protect health, environment, and natural resources from persistent and mobile chemicals Preventing Recalcitrant Organic Mobile Industrial chemicalS for Circular Economy in the soil-sediment-water System Start date of the project: 1st November 2021 Duration: 42 months Authors: Veronika Zhiteneva (KWB), Malte Zamzow (KWB), Antoine Daurat (KWB), Mireia Mesas (EURECAT), Hèctor de Buen (EURECAT) Lead Beneficiary: KWB Type of delivery: DEM Dissemination Level: PU Filename and version: PROMISCES_D2.5_Model-HHREA (version 1) SharePoint: www.promisces.eu Due date: 30 April 2024 (M30) Date of revision: 30 April 2025 (M42) D2.5 – Open source model for probabilistic human health risk assessment © European Union, 2025 No third-party textual or artistic material included on the publication without the copyright holder’s prior consent to further dissemination by other third parties. Reproduction is authorized provided the source is acknowledged Disclaimer The information and views set out in this report are those of the author(s) and do not necessarily reflect the official opinion of the European Union. Neither the European Union institutions and bodies nor any person acting on their behalf may be held responsible for the use which may be made of the information contained therein. D2.5Open source model for probabilistic human health risk assessment 2 Document History This document has been through the following revisions: Version date Author Description 0.1 02.04.2024 Veronika Zhiteneva, Malte Zamzow, Mireia Mesas, Hèctor de Buen Draft ready for reviewer check 0.2 23.04.2024 Veronika Zhiteneva, Malte Zamzow Incorporated reviewer comments 0.3 25.04.2024 Julie Lions Version 0.3 for submission 0.4 16.09.2024 Veronika Zhiteneva, Malte Zamzow History of changes: P6: Added list of figures, P7: Added list of tables, P30: updated delivery dates in Table 5, split CS6 technologies into 2 separate technologies, P47: added schedule of next steps and delivery of experimental data and interaction with DSF (reviewer request) to future work section 0.5 10.04.2025 Veronika Zhiteneva, Malte Zamzow, Antoine Daurat, Mireia Mesas, Hèctor de Buen Final version for review (part I) 0.6 15.04.2025 Veronika Zhiteneva, Malte Zamzow, Antoine Daurat, Mireia Mesas, Hèctor de Buen Incorporate reviewer comments and provided complete final version for review 0.7 28.04.2025 Veronika Zhiteneva, Malte Zamzow, Antoine Daurat, Mireia Mesas, Hèctor de Buen Incorporate reviewer comments and submitted to WPL 0.8 29.04.2025 Veronika Zhiteneva Final version 0.9 29.04.2025 Floriane Sermondadaz Quality control 1.0 29.04.2025 Julie Lions Final version for submission D2.5Open source model for probabilistic human health risk assessment 3 Authorisation Distribution This document has been distributed to: Authorisation Name Status Date Review Martine Bakker Completed 12.04.2024 (Version 0.3) 11.04.2025 Version 0.5 17.04.2025 (Version 0.6) Validation Hans Groot, Petra Hogervorst WP Leaders 12.04.2024 (Version 0.3) 19.09.2024 (Version 0.4) 28.04.2025 (Version 0.7) Quality Control Julie Lions Floriane Sermondadaz Project coordinator Administrative and financial manager 25.04.2024 (Version 0.3) 20.09.2024 (Version 0.4) 29.04.2025 (Version 0.8 and 0.9) Approval Julie Lions Project coordinator 25.04.2024 (version 0.3) 20.09.2024 (version 0.4) 29.04.2025 Version 0.8 Name Title Version issued Date of issue BRGM, EURECAT, KWB, DELTARES, RIVM WP2 partners Version 0.3 25.04.2024 BRGM, EURECAT, KWB, DELTARES, RIVM WP2 partners Version 1 29.04.2025 D2.5Open source model for probabilistic human health risk assessment 4 Executive Summary A risk-based human health exposure assessment (HHEA) model was developed to evaluate the exposure for humans in 4 circular economy (CE) routes investigated in 6 of the 7 case studies in the project PROMISCES. The HHEA is a probabilistic tool evaluating the risk posed to human health. The HHEA was applied to the following routes: 1) semi-closed drinking water cycle; 2) groundwater remediation; 3) water reuse for agricultural irrigation; and 4) nutrient recovery. Each of these exposure routes results in a product – drinking water or lettuce – which can be consumed by humans. For some routes, the exposure is purely theoretical, while for others, the entire process chain is investigated in the PROMISCES case study. The HHEA is built on Bayesian principles and includes Bayesian updating, which enables assessment of risk under conditions of low data availability and high uncertainty. This is particularly useful for evaluation of substances such as PFAS and other industrial persistent, mobile and potentially toxic (iPMT) substances, the removal of which in treatment processes is not yet well studied in literature. The deliverable explains the different treatments, environmental matrices, and substances which were the focus of the initial assessment. It describes the construction of the HHEA model, with explanations of how different data types – literature data, site specific data, and modelled data – are used to update the prior probability of the removal factor for substances in a process. It also describes how non-technical processes, such as mixing or evaporation, have been included into the treatment trains evaluated. Finally, individual reference quotients for the substances are established, which are used to assess the relative risk of the final concentrations in the products which could be consumed by humans. In the semi-closed water cycle route, six priority substances (PFOA, PFNA, PFDA, PFPeS, PFOS, and TFMSA) were identified with risk quotients (RQs) exceeding 1 in the baseline. Scenario analysis revealed that river dilution had the most significant effect on reducing compound concentrations. Additionally, granular activated carbon (GAC) proved effective in lowering risks, regardless of the background contamination level of the water body from which the water for GAC treatment was taken. This finding highlights the importance of both hydrological context and advanced treatment technologies in exposure mitigation. The model’s sensitivity to data type and quality was demonstrated using PFOS, TFMSA, and carbamazepine in the semi-closed water cycle route. While PFOS had abundant but inconsistent literature data, TFMSA and carbamazepine relied on measured values, leading to contrasting outcomes: low removal for TFMSA and high for carbamazepine, particularly in GAC treatment. This underscores the critical role of both measured and literature data in refining risk predictions and reinforces the need for reliable, transferable datasets to reduce uncertainty in PFAS exposure assessments. For the groundwater remediation route, all compounds showed RQs far above 1 due to assumed high starting concentrations (5 µg/L) and limited removal data. Unlike the other routes, risk here was primarily driven by the initial concentration relative to health-based thresholds. The analysis showed that the more confident we are in high removal efficiency (e.g., through repeated measurements), the higher the allowable starting concentration can be. This emphasizes that model uncertainty D2.5Open source model for probabilistic human health risk assessment 5 around treatment performance can have a larger influence on risk outcomes than the choice of treatment technology itself. In the water reuse route, six substances consistently presented RQs above 1 across all percentiles. Scenario analysis showed that for PFOA, inlet concentration played a greater role in risk than treatment variability, while for PFOS, even improved treatment could not bring RQ below 1 due to its reference value. Conversely, in the nutrient recovery route, all seven PFAS compounds had RQs at or below 1 across all percentiles, mainly due to effective removal during membrane filtration, suggesting that risks from this route are comparatively low under current conditions. A key strength of the HHEA model lies in how it handles epistemic uncertainty—uncertainty stemming from limited or variable data. The model weighs negative findings more heavily in shaping the likelihood and posterior distributions, encouraging precaution in the absence of consistent highremoval evidence. When site-specific measurements are available, the model integrates them randomly with concentration values to avoid under or over estimating risk, supporting more riskbased decision-making. Overall, the HHEA model can be successfully implemented for a variety of available data quantities and end goals, which enables flexible exposure assessment and simplified visualization of which compounds should be prioritized for remediation. The model is available at https://github.com/KWB-R/promisces.hhea. D2.5Open source model for probabilistic human health risk assessment 6 Table of contents 1. Introduction ................................................................................................................................. 11 2. Application in PROMISCES ........................................................................................................... 12 Model approach ................................................................................................................... 12 PROMISCES exposure routes and initial compounds of interest ........................................ 12 Matrix implementation in the HHEA database .................................................................... 14 Treatment implementation in the HHEA database ............................................................. 14 3. Model description ........................................................................................................................ 16 Bayesian background ........................................................................................................... 16 Standard removal process ................................................................................................... 17 4.2.1 The prior probability ........................................................................................................... 17 4.2.2 Updating the unknown ................................................................................................ 18 4.2.3 Combining input concentration and removal factors ................................................. 21 Mixing and demixing processes ........................................................................................... 22 Sewage sludge as an output of wastewater treatment ...................................................... 23 Producing a reference quotient ........................................................................................... 23 Model implementation (GitHub) ......................................................................................... 26 4. Model Application ........................................................................................................................ 27 Semi-closed water cycle ....................................................................................................... 29 4.1.1. PFOS, TFMSA and carbamazepine: HHEA result interpretation (baseline scenario) .. 32 4.1.2. All substances: baseline scenario and scenario 1 ........................................................ 35 4.1.3. Priority substances: comparing different scenarios .................................................... 37 4.1.4. Priority substances: backwards estimation of tolerable starting concentration ........ 39 Groundwater remediation ................................................................................................... 40 Water reuse ......................................................................................................................... 44 Nutrient recovery ................................................................................................................. 52 Code availability ................................................................................................................... 62 5. Discussion and conclusions .......................................................................................................... 63 6. Literature ..................................................................................................................................... 65 Annexes ................................................................................................................................................ 67 Annex I. Water reuse simulation outputs ............................................................................................ 67 Annex II. Nutrient recovery simulation outputs .................................................................................. 96 D2.5Open source model for probabilistic human health risk assessment 7 List of figures Figure 1: Semi-closed water cycle exposure route (CS 1 and 2 in PROMISCES). ................................. 13 Figure 2: Groundwater remediation exposure route (CS 6 and 7 in PROMISCES). ............................. 13 Figure 3: Water reuse for agricultural irrigation exposure route (CS 3 in PROMISCES). ..................... 13 Figure 4:Nutrient recovery exposure route (CS 4 in PROMISCES). ...................................................... 13 Figure 5: Depiction of how parent nodes are connected to a child node via arcs in a Bayesian Network. .............................................................................................................................................................. 16 Figure 6: Beta distribution as the prior for the removal factor. .......................................................... 18 Figure 7: Schematic overview of how literature data are used (fRM: removal factor, n: number of all available removal factors) to derive the likelihood distribution. ........................................................ 19 Figure 8: Schematic overview of how case study data are used to derive the likelihood distribution (fRM,i: removal factor based on measurement i; fRM,con: overall removal factor including the conservative starting point; n: number of samples). ........................................................................... 21 Figure 9: Semi-closed water cycle exposure route. ............................................................................. 29 Figure 10: Spider plot of the baseline scenario for PFOS, TFMSA, and carbamazepine. .................... 33 Figure 11: Violin plots of posterior removal factor distributions for PFOS (red), TFMSA (orange) and carbamazepine (green) along the treatment processes of the baseline scenario. Treatment IDs were defined in Table 2. ............................................................................................................................... 34 Figure 12: Removal factor distribution of drinking water treatment activated carbon for A: PFOS, B: TFMSA and C: carbamazepine ............................................................................................................. 34 Figure 13: Removal factor distributions of primary and secondary wastewater treatment for A: PFOS and B: TFMSA and carbamazepine. ..................................................................................................... 35 Figure 14: Spider plot of the water cycle baseline scenario (GAC as final drinking water treatment) .............................................................................................................................................................. 36 Figure 15: Spider plot of the water cycle scenario 1 (IX as final drinking water treatment) .............. 37 Figure 16: Spider plots of all water cycle scenarios for the priority substances ................................. 38 Figure 17: Spider plot of PFOA for different starting concentrations (Baseline scenario) .................. 39 Figure 18: Groundwater remediation exposure route. ....................................................................... 40 Figure 19: Spider plots of the groundwater remediation (A: persulfate oxidation, B: ultrasonic cavitation) baseline scenario ............................................................................................................... 42 Figure 20: Water reuse exposure route. .............................................................................................. 44 Figure 21: Modelled removal factors and concentration profile for PFOA in the water reuse route. 46 Figure 22: Spider plot for all 14 substances of water reuse route. ..................................................... 47 Figure 23: Modelled PFOA concentration profiles in the water reuse exposure route for the baseline scenario (Std) and4 adapted scenarios. The reference value is 6 ng/L PFOA. .................................... 50 Figure 24: Modelled PFOA RQ for the water reuse exposure route for the baseline scenario (Std) and 4 adapted scenarios. ............................................................................................................................ 51 D2.5Open source model for probabilistic human health risk assessment 8 Figure 25: Nutrient recovery exposure route. ..................................................................................... 52 Figure 26: Modelled removal factors and concentration profile of nutrient recovery route for PFOA. .............................................................................................................................................................. 55 Figure 27: Spider plot for all 7 PFAS in nutrient recovery route.......................................................... 56 Figure 28: Modelled concentration profiles of nutrient recovery route – scenarios 1 to 4. Reference value is estimated at 6 ng/L PFOA. ...................................................................................................... 61 Figure 29: Modelled reference quotient diagram along the nutrient recovery route for PFOA in 4 different scenarios. .............................................................................................................................. 62 D2.5Open source model for probabilistic human health risk assessment 15 Table 2: Treatment processes used in the HHEA. ID Name Input Matrix Output Matrix wwt1 Primary wastewater treatment (mechanical and physical treatment) rww, iww, hww, stw, lww no_change wwt2 Secondary wastewater treatment rww, iww, hww, stw, lww tww wwtt Combination of primary and secondary wastewater treatment rww, iww, hww, stw, lww tww wwsl Sludge production after primary and secondary wastewater treatment rww, hww sdg wwco Tertiary wastewater treatment: Coagulation and filtration tww, iww tww wwuf Tertiary wastewater treatment: Ultrafiltration tww, iww tww wwnf Tertiary wastewater treatment: Nanofiltration tww, iww tww wwro Tertiary wastewater treatment: Reverse Osmosis tww, iww tww wwel Tertiary wastewater treatment: Electro-oxidation (e-peroxone) tww, iww tww wwmb Tertiary wastewater treatment: Membrane bioreactor tww, iww tww grpf Groundwater treatment: Persulfate-based in situ chemical oxidation - (n)ZVI + Fe(VI) grw no_change gruc Groundwater treatment: ultrasonic cavitation grw no_change gren Groundwater treatment: catalysis by extra cellular ligninolytic enzymes grw grw npbk Natural Process: Bank filtration suw bfw wetl Additional treatment: Constructed wetlands rww, iww, hww, stw, tww, lww tww dwae Drinking water treatment: aeration suw, bfw, grw, drw, tww, stw drw dwrf Drinking water treatment: rapid filtration suw, bfw, grw, drw, tww, stw drw dwac Drinking water treatment: granular activated carbon suw, bfw, grw, drw, tww, stw drw dwex Drinking water treatment: anion exchange suw, bfw, grw, drw, tww, stw drw dilsw Dilution by surface water tww, stw suw dilww Dilution by municipal wastewater iww, lww, rww, tww rww dilgw Dilution by groundwater bfw, pow grw dilrw Dilution by soil irrigation water (rain water) suw, tww, stw, sdg pow dilpr Dilution by process water (from minor secondary stream) tww, iww, hww, rww, lww, drw, grw no_change sepev Separation due to evaporation suw, pow no_change NaturalProcess: Degradation in groundwater (biotic and abiotic) D2.5Open source model for probabilistic human health risk assessment 16 3. Model description Bayesian background The HHEA was originally foreseen as a probabilistic Bayesian network (BN). BNs are probabilistic graphical models in which variables (nodes) are connected to each other via arcs (arrows) going from parent to child nodes (Martinez-Megias et al., 2023). An example can be seen in Figure 5. Figure 5: Depiction of how parent nodes are connected to a child node via arcs in a Bayesian Network. Each node has a set of possible states (parameters, values, etc.) that occur with a certain probability. The probabilities of the states are known as beliefs. The nodes are described using probability distributions, which can be fit to existing information such as expert knowledge, experimental data, literature data, or a mix of data sources. This fitting of some type of existing information to a probability distribution produces a prior probability. The arcs represent conditional probability tables, which define the relationship between the parent and child nodes and how they change. Using the belief and the established network, Bayesian inferencing using Bayes’ theorem can be used to calculate the risk at the end of the exposure route. Bayes’ theorem (Equation 1) calculates the posterior probability ( of the child A given the parent B, given the prior probability of the child, the marginal probability of the parent , and the likelihood of the child given a fixed state of the parent. 𝑃𝑃(𝐴𝐴|𝐵𝐵)=𝑃𝑃(𝐵𝐵|𝐴𝐴)∗𝑃𝑃(𝐴𝐴) 𝑃𝑃(𝐵𝐵) Equation 1 Using this equation, BNs can show different states of nodes as they are updated with more data (belief updating) and different types of data. BNs are easy to visualize and understand by nontechnical audiences. They are also suitable for small and incomplete data sets (Uusitalo, 2007), which makes them applicable for conducting risk assessments of iPMT and PFAS in CE routes. In the process of tool development, a Bayesian network was foregone for a probabilistic tool with Bayesian updating. The benefit of a Bayesian network, namely that it allows backwards inferencing, was difficult to implement in this tool due to the low data availability, and lack of experimental data in comparison to literature data. Additionally, a Bayesian network requires data to be discretized into categories, which would further reduce the resolution with which results could be analysed for the CE routes explored herein. Therefore, it was decided to keep the tool as a probabilistic model with Bayesian updating. However, upgrading to a Bayesian network could be done in the future. D2.5Open source model for probabilistic human health risk assessment 17 Standard removal process For most of the technical treatment processes, the output concentration can be estimated by using an input concentration and a removal factor (RF), as seen in Equation 2. This is the standard procedure in the HHEA model. 𝑐𝑐𝑜𝑜𝑜𝑜𝑜𝑜 =𝑐𝑐𝑖𝑖𝑖𝑖 (1−𝑅𝑅𝑅𝑅) Equation 2 The input concentration is either a single value, a uniform distribution by a range defined as starting concentration, or the result of the simulation of previous processes. The removal factor distribution is built by prior and likelihood distributions. The procedure is described in detail in the following sections. 4.2.1 The prior probability The overall goal of the risk-based HHEA is not to estimate the most probable substance concentration along the exposure route, but to describe the final risk of exceeding a reference value. The risk is the result of the impact of a negative effect and its likelihood to occur, and is described by Equation 3. 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 = 𝐿𝐿𝑅𝑅𝑅𝑅𝐿𝐿𝐿𝐿𝑅𝑅ℎ𝑜𝑜𝑜𝑜𝑜𝑜 𝑥𝑥 𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝑐𝑐𝐼𝐼 Equation 3 The knowledge about both parts of the aforementioned equation must be considered. An unacceptable risk occurs when: • given a certain impact, the likelihood is too high or unknown • given a certain likelihood, the impact is too high or unknown The prior distribution determines how much information is needed to prove low risk. Therefore, it is the most important assumption for assessing situations where knowledge or data is insufficient. For the PROMISCES HHEA, the prior for the removal within one process was required to meet the following criteria: 1. the distribution needs to be symmetric - high and low removal must be equally likely – to avoid an initial bias. 2. Both options - complete removal and no removal - must be included in the distribution. Furthermore, regarding the process chain, the following must be true: 1. For several subsequent processes with unknown removal, the outcome “no removal at all” must not be excluded. 2. Each additional process with unknown but potential removal must result in a lower final risk. The most intuitive choice for the prior would be a uniform distribution, which represents every removal factor as equally likely if nothing is known about the removal factor. However, drawing from a uniform distribution between 0 and 1 leads to very few values close to 0% removal. This would mean that for several subsequent processes with no information on the removal factor, there would D2.5Open source model for probabilistic human health risk assessment 18 be no scenario with repeatedly 0% removal in every process step, unless the number of samples drawn from the distributions is (inefficiently) large. Thus, a worst case scenario cannot be modelled by a uniform distribution. Instead, a convex shaped beta distribution was chosen as prior for the removal factors (Figure 6). The beta distribution is described by two parameters, alpha and beta, and is only defined for values between 0 and 1. It is symmetric when both of its parameters are equal (criterion 1). For alpha and beta below 1, the density increases towards both 0 and 1 (criterion 2). This distribution favours low or high removal and reduces the weights of intermediate values. For removal factors, this means that best case and worst case removal scenarios are favoured by default. This way, even long treatment chains contain the option “no removal at all” with a reasonable amount of samples drawn (criterion 3), while at the same time, the average risk decreases with every process step (criterion 4). Figure 6: Beta distribution as the prior for the removal factor. 4.2.2 Updating the unknown According to Bayes’ theorem (Equation 1), the posterior distribution describes the final probability of one process. It is a combination of the prior assumption and the likelihood based on observations. For the HHEA, the prior can be updated either by literature and/or by on-site sample data. All removal factors provided for the literature or site based likelihood distribution must be integers and expressed as a percentage (%). Sometimes, only qualitative statements about the substance removal are available in literature. The qualitative statement “no removal observed” and “no significant removal observed” are added as 0% and 5% removal, respectively. The posterior of a removal factor is proportional to the prior, multiplied by the likelihood based on literature data and the likelihood based on measured data. Updating the prior with literature data Literature data from a study is only transferable to a limited extent to other locations. Varying environmental conditions or technical specifications have a large impact on the removal factor. In most cases, the potential variation between two identical processes at two different locations is larger than the variation within one location. Thus, the suitability of a removal factor from an external data source for the current assessment cannot be estimated by the standard deviation of the factor D2.5Open source model for probabilistic human health risk assessment 19 or other statistical dispersion parameters. In other words: regardless of the spread of removal factors at two different sites, the mean or median at one site can still be informative. For that reason, only one value per site, such as a location parameter like the mean or median, is considered in the literature likelihood (Figure 7). In reviews or meta studies (i.e. a study evaluating the results of other studies), single results are often aggregated into ranges. Here, for the reason explained above, instead of using the range of a meta study, it is necessary to list the results of single studies separately. In addition to the literature data, a conservative starting point (the conservative removal factor) is included to account for the amount of knowledge involved. It can be described as the mean of all included removal factors and 0.1% removal. If the number of literature data is low, the conservative removal factor is significantly lower than the actual mean value of all removal factors found in the literature. As the number of data increases, the conservative removal factor gets closer to the actual mean of the literature data. The more data available, the less conservative the likelihood becomes. Three examples are listed in Table 3 to explain the approach. Figure 7: Schematic overview of how literature data are used (fRM: removal factor, n: number of all available removal factors) to derive the likelihood distribution. D2.5Open source model for probabilistic human health risk assessment 20 Table 3: Example of how likelihood and posterior distributions are affected based on different types of literature data evaluated. Example 1 study 80% removal 5 studies 80% removal 1 study 20% removal Figure Explanation • High impact of the conservative removal factor on the likelihood • Broad likelihood distribution • The prior dominates the shape of the posterior, but it’s no longer symmetric • Shift towards higher probability of complete removal • Low impact of the conservative removal factor on the likelihood • Very strong evidence that the removal factor is indeed somewhere around 80% • The likelihood dominates the shape of the posterior • One study with 20% removal has a stronger effect on the posterior than one study with 80% removal • Risk-based approach: more evidence is needed to verify a high removal Updating the prior with site specific data The prior distribution can also be updated with experimental data generated directly in the particular exposure route in PROMISCES. This data is understandably most suitable for the specific CE exposure route. If it is available, the prior is updated in two stages: first by literature data and then by case study data. However, a decision on whether to use only case study data if literature data insufficiently reflect the process can also be taken. This can occur when a case study process is optimized for the removal of a specific substance, while literature data is based on a general process setup. Similar to the procedure for updating using literature data explained above, a likelihood for the measured data must also be derived. Instead of using aggregated values, every removal factor measured in the case study will be included in the likelihood, which gives more weight to the site specific results (Figure 8). If literature data is included, the posterior distribution is then calculated as the product of the prior distribution and two likelihood distributions, one based on literature data, and based on case study data. Concentrations measured below the limit of detection (LOD) can be considered as complete removal, i.e. a removal factor of 100%. The conservative starting point requires a large amount of measurements below LOD in order to have a high impact on the posterior distribution. D2.5Open source model for probabilistic human health risk assessment 21 Figure 8: Schematic overview of how case study data are used to derive the likelihood distribution (fRM,i: removal factor based on measurement i; fRM,con: overall removal factor including the conservative starting point; n: number of samples). Updating the prior with model data Many natural and technical processes can be described with various types of models. A major advantage of modelling is that input variables can be quickly adjusted and new results can be generated. The number of possible model results is therefore easily many times higher than measurement results. On the other hand, measurement results often represent reality better than models. One example is a model that describes the removal of a substance by GAC, where the removal efficiency is dependent on the bed volume and the DOC content of the water, among other things. It would be possible to calculate numerous results by varying the input variables. Feeding all the model results into the HHEA model would dominate the posterior removal factor distribution due to the amount of data available. Instead, calculating only two results is suggested: i) based on average conditions; and ii) adjusted to the specific conditions of the case study. Using this approach, one model will have exactly the same weight as one literature study or one on-site measurement. 4.2.3 Combining input concentration and removal factors The HHEA combines two important aspects from the literature and the experiments in PROMISCES: concentration and removal percentages of substances in the different steps of the exposure routes. This section explains how these two aspects are utilized in the HHEA. The input concentration of a substance and the removal factor of a process can either be paired randomly, or sorted so that the lowest removal factor is paired with the highest initial concentration and vice versa (worst case and best case scenarios). Although the sorting (worst case and best case) seems to fit to the idea of conservative risk assessment, it may lead to an unjustifiable overestimation of risk. Therefore, two different methods for data pairing were implemented in the HHEA model: 1. If the difference in removal factors is expected to average out over time, removal factors and input concentrations are combined randomly. Example case 1: Raw wastewater concentrations are paired with the removal in a WWTP over 1 year – as it’s highly unlikely that the highest influent concentration and the lowest WWTP removal would be consistently paired over the course of the whole year, random pairing is allowed. D2.5Open source model for probabilistic human health risk assessment 22 2. Removal factors and input concentrations are paired in a best case and worst case manner. Example case 2: Data from two different treatment plants show different removals (30% removal and 60% removal), for which some site specific variability is known (i.e. different ozone doses). If the removal factor would be applied randomly, this systematic difference would be ignored. Therefore, input concentrations and removal factors are paired so that the best case and worst case scenarios are assessed. In general, data that is gathered in the case studies can be assumed to average out over time and are paired randomly (example case 1), while literature data can have systematic site specific deviation and are assessed for best case and worst case scenarios (example case 2). During the simulation, the dominant input distribution is defined for each posterior distribution. If this is the site specific case study data, method 1 is applied. The site specific data is assumed to be dominant if the mode of the likelihood is less than 5% away from the mode of the resulting posterior. The 5% was selected based on the frequently used limit of 5% as the significance level: if the difference between the two modes is less than 5%, then it is not significant. Mixing and demixing processes In several processes, different types of water are mixed (e.g. effluent into surface water, bank filtrate into groundwater) or separated (e.g. evaporation from soil pore water). The output concentration of a mixture of two waters is defined in Equation 4: 𝑐𝑐𝑜𝑜𝑜𝑜𝑜𝑜 =𝑐𝑐𝑖𝑖𝑖𝑖𝑥𝑥𝑖𝑖𝑖𝑖 +𝑐𝑐2𝑥𝑥2 Equation 4 where is the concentration in the input, is the concentration in the final water, is the according proportion in the input, and is the proportion being mixed. The sum of and is equal to 1. Equation 4 is also the basis for the separation process, where a proportion of the input water(x2) leaves the system with concentration. The remaining substance is thus concentrated in the output. This can be described by Equation 5: 𝑐𝑐𝑜𝑜𝑜𝑜𝑜𝑜 =𝑐𝑐𝑖𝑖𝑖𝑖 −𝑐𝑐2𝑥𝑥2 𝑥𝑥𝑜𝑜𝑜𝑜𝑜𝑜 Equation 5 All mixing and demixing data are very site specific and cannot be substituted with literature data. Conversely, it is assumed that mixing or demixing always averages out, and process parameters are randomly combined with the process input concentration. Both proportions sum up to one, so only one proportion x2 and optionally a corresponding concentration c2 need to be provided. For standard removal, the mixing and demixing are applied in a probabilistic way. Data can be entered either as mean and standard deviation, or as minimum and maximum proportion or concentration. Regardless of the type of values provided, a normal distribution is created. The range between a minimum and maximum value is therefore interpreted as a 95% interval of a normal distribution. The normal distribution of the water proportion and the concentration is truncated at 0 and 1, and at 0 ng/L. For the separation process, the distribution is also cut off at the highest possible concentration, which corresponds to the water mixture as it enters the process. D2.5Open source model for probabilistic human health risk assessment 23 Sewage sludge as an output of wastewater treatment Substances can be removed from the water matrix during the primary and secondary wastewater treatment by • mechanical elimination; • biological degradation; or • sorption to sludge followed by sludge separation. For PFAS and other iPMTs, mechanical removal and biodegradation play a minor role. If any removal occurs, it is usually due to sorption to sludge. Sludge thickening and dewatering can be seen as a separation process, in which treated wastewater is separated from a matrix containing a higher proportion of solids. Thus, it is also based on Equation 6, where the concentration of the separated liquid phase is equal to the secondary wastewater treatment effluent concentration. The distribution of the concentration in the effluent can be obtained by applying the standard removal process for primary and secondary wastewater treatment (wwtt). 𝑐𝑐𝑠𝑠𝑠𝑠𝑜𝑜𝑠𝑠𝑠𝑠𝑠𝑠 =𝑐𝑐𝑖𝑖𝑖𝑖 −(1−𝑅𝑅𝑅𝑅)𝑐𝑐𝑖𝑖𝑖𝑖�1−𝑥𝑥𝑠𝑠𝑠𝑠𝑜𝑜𝑠𝑠𝑠𝑠𝑠𝑠� 𝑥𝑥𝑠𝑠𝑠𝑠𝑜𝑜𝑠𝑠𝑠𝑠𝑠𝑠 Equation 6 A sludge volume of 10% of the inflow volume of raw wastewater is assumed as an average. A standard deviation of 2% is added to account for variations between different treatment plants. For degradable substances or substances removed to a significant extent by mechanical wastewater treatment, calculating sludge concentration within the HHEA approach is much more complicated and was therefore not implemented. Producing a reference quotient The final output concentrations are compared to a reference value stemming from international or national guidelines. Equation 7 was used derive a reference quotient (RQ): 𝑅𝑅𝐿𝐿𝑅𝑅𝐿𝐿𝑅𝑅𝐿𝐿𝑅𝑅𝑐𝑐𝐿𝐿 𝑞𝑞𝑞𝑞𝑜𝑜𝐼𝐼𝑅𝑅𝐿𝐿𝑅𝑅𝐼𝐼 (𝑅𝑅𝑅𝑅) = 𝑐𝑐𝑜𝑜𝑜𝑜𝑜𝑜 𝑅𝑅𝐿𝐿𝑅𝑅𝐿𝐿𝑅𝑅𝐿𝐿𝑅𝑅𝑐𝑐𝐿𝐿 𝑣𝑣𝐼𝐼𝐿𝐿𝑞𝑞𝐿𝐿 Equation 7 For exposure routes ending in drinking water, the latest European drinking water directive (Directive 2020/2184) value for PFAS-20 of 100 ng/L was used to derive a reference value for each individual PFAS. 100 ng/L were first divided by 20 which results in an effective single substance fraction of 5 ng/L. Relative potency factors (RPF), which quantify the effect of a single substance compared to PFOA, which has the RPF = 1, were used. By dividing the PFOA fraction of 5 ng/L by the RPF, an effectively equivalent reference value is obtained for the individual substances. The RPF and the derived reference values are given in Table 4 and come from (RIVM., 2023). Use of the RPFs enables the HHEA to consider in vivo endpoints of the individual PFAS and further substances which are not part of the PFAS-20 sum of the directive can be included in the assessment. It is important to note that the HHEA is only applicable for liquid media. A concentration in other matrices, like plant tissue, cannot be simulated with this approach. Instead of calculating the whole pathway from irrigation water or sludge to human crop consumption (relevant for the nutrient D2.5Open source model for probabilistic human health risk assessment 24 recovery and water reuse for agricultural irrigation exposure routes), a reference value for soil pore water was calculated backwards, starting with the tolerable daily intake (TDI) was available for PFOA (Schrenk et al., 2020) and for dibuthyl phthalate (DBP) (Silano V, 2019). To obtain the TDI of individual substances, the same procedure was applied as for obtaining the drinking water reference values. The TDI was calculated for a person weighing 70 kg. It was further assumed that the food under consideration should contribute only 10% to the TDI value (e.g. a safety factor for other sources). To calculate the pore water concentration , the average daily consumption of vegetables (0.26 kg according to (Ministerio de Medio Ambiente, 2007)) and the bioconcentration factor (BCF) from the soil pore water to the plant foliage were also required (Figure 10). TDI and BCF are substance-specific values (in silico endpoints), whereas the remaining values in Equation 8 were applicable for substances considered. 𝑐𝑐𝑝𝑝𝑝𝑝 =𝑇𝑇𝑇𝑇𝐼𝐼�𝑅𝑅𝑛𝑛 𝑅𝑅𝑛𝑛∗𝑜𝑜�×70[𝑅𝑅𝑛𝑛] 0.26 �𝑅𝑅𝑛𝑛 𝑜𝑜�×𝐵𝐵𝐵𝐵𝑅𝑅�𝐿𝐿 𝑅𝑅𝑛𝑛�×10% Equation 8 If a TDI was not available, relative potency factors (RPF) were used when available. For the PFAS compounds, RPF list by (RIVM., 2023) was considered. These RPFs quantify toxicity relative to PFOA (index compound), converting mixed PFAS exposures into PFOA equivalents and assuming dose additivity for shared toxicological endpoints. There was no TDI or RPF for diethyl phthalate, carbamazepine and diuron. In such cases, pore water reference concentration was calculated according to US Environmental Protection Agency methodology (EPA., 1989) with exposure parameters established by Spanish Law (Estado., 2003; Ministerio de Medio Ambiente, 2007)) and toxicological data from several sources (Health., 2013; RAIS., 2025; TCEQ., 2023). It should be noted that, although reference values were derived for this analysis, this was not one of the main objectives of the intended work. For substances for which there were no published reference values, other regulatory values could have been considered, such as those from the EU Drinking Water Directive 2020/2184 (Commission., 2020). However, we chose to derive them to avoid an extremely conservative approach and to better illustrate the correct functioning of the developed tool with more realistic thresholds. BCF values were obtained from (Felizeter, 2020) for most of the PFAS. If BCF value was not available for a substance, it was estimated that an average of 150 L of irrigation water is needed to obtain 1 kg of lettuce. Reference values used for the different water matrices evaluated in the HHEA can be seen in Table 4. D2.5Open source model for probabilistic human health risk assessment 31 Table 7: Removal factors derived from experimental data of CS1 Substance Removal factors in % GAC (Bed volumes 10,000 – 20,000) IX (Bed volumes 30,000 – 40,000) (only for 6 priority substances) PFBA 8, 7, 5, 8, 4, 4, 5 PFPeA 6, 4, 4, 15, 10, 15, 11 PFHxA 7, 4, 9, 26, 25, 27, 32 PFHpA 17, 12, 19, 25, 20, 29, 31 PFOA 27, 21, 21, 39, 23, 40, 30 53, 92 PFNA 30, 17, 23, 35, 18, 59, 39 67,98 PFDA 33, 42, 34, 26, -8, 74, 34 77,99 PFBS -9, 6, -12, 33, 26, 32, 29 PFPeS 31, 21, 30, 42, 36, 41, 38 96, 100 PFHxS 31, 23, 31, 43, 36, 51, 39 PFOS 33, 27, 28, 42, 18, 69, 34 98, 100 TFMSA 6, 8, 5, 15, 7, 3, -3, 4 68, 69, 48, 24 Benzotriazole 88, 89, 89, 85, 89, 88, 78, 79 Carbamazepine 99, 98, 99, 98, 97, 94, 99, 99 The WWTP effluent discharges into surface water and is thereby diluted. In the baseline scenario, clean surface water (no background concentrations) from a low-flow river is assumed. One part of treated wastewater (33.3%) is mixed with two parts of clean surface water (66.6%) on average. This ratio changes to 5% treated wastewater and 95% surface water in the scenarios that contain a large river. Since no site-specific data can be used for those generic assumptions, "contamination" was defined as a substance concentration twice as high as the reference value. A variation based on 20% standard deviation was integrated into the simulation. The second dilution process of the treatment train remained constant for all scenarios. Following the bank filtration process, 75% of bank filtrate is assumed to be mixed with 25% clean groundwater (Table 8). D2.5Open source model for probabilistic human health risk assessment 32 Table 8: Dilution characteristics of the water cycle baseline scenario Process Average proportion of diluting water Standard deviation of diluting water Average pollutant concentration Standard deviation of pollutant concentration Dilution in a small clean river (Baseline Scenario, Scenario 1) 0.666 0.067 0 0 Dilution in a large clean river (Scenario 2) 0.95 0.025 0 0 Dilution in a small contaminated river (Scenario 4) 0.666 0.067 Twice the reference value of the substance 20% of the average concentration Dilution in a large contaminated river (Scenario 3) 0.95 0.025 Twice the reference value of the substance 20% of the average concentration Dilution by groundwater 0.25 0.05 0 0 4.1.1. PFOS, TFMSA and carbamazepine: HHEA result interpretation (baseline scenario) The spider plot in Figure 10 shows the result of the baseline scenario for PFOS, TFMSA and carbamazepine. It compares the likelihood of exceeding a predefined reference value by displaying the upper part (50th to 99th percentiles) of the RQ distribution. The 75th percentiles of PFOS and TFMSA reference quotient distributions are more than tenfold higher than the reference value. Thus, there is a high probability that the reference value will be significantly exceeded. For PFOS, even the 50th percentile of the distribution is already well above the reference value. For carbamazepine, the HHEA model does not return a reference value above 1. In the subsequent paragraphs, the interim results of the baseline scenario for the high-risk substances PFOS and TFMSA and the low-risk substance carbamazepine were compared to explain how the HHEA model works and to interpret the results. These compounds were chosen due to their different RQs and different levels knowledge. D2.5Open source model for probabilistic human health risk assessment 33 Figure 10: Spider plot of the baseline scenario for PFOS, TFMSA, and carbamazepine. One main difference between the simulations of PFOS, TFMSA, and carbamazepine removal is the availability of literature data. While many data points for PFOS removal are listed in the HHEA data base, information about TFMSA and carbamazepine removal is missing. Despite this, the results for both substances differ significantly, which can be explained through the presence of case study data added. For all treatment processes (except GAC), the applied removal factor distribution is very similar for TFMSA and carbamazepine (Figure 11). However, the measured data for TFMSA and carbamazepine removal by GAC treatment indicate large differences in the removal efficiency (Table 7). D2.5Open source model for probabilistic human health risk assessment 34 Figure 11: Violin plots of posterior removal factor distributions for PFOS (red), TFMSA (orange) and carbamazepine (green) along the treatment processes of the baseline scenario. Treatment IDs were defined in Table 2. Figure 12 shows the impact of literature and measured data on the posterior distribution for GAC. The high standard deviation of PFOS literature data leads to a flat likelihood distribution. However, there is a noticeable peak of the literature – likelihood distribution at around 80%. On the contrary, the rapid small scale column test (RSSCT) results indicate a lower removal by GAC (Rückbeil et al., 2025). The posterior distribution is in between both likelihood distributions. No literature data for TFMSA or carbamazepine was collected. The data from the experiments are very contradictory, leading to almost certainly low removal for TFMSA and high removal for carbamazepine. While TFMSA is poorly removed, carbamazepine is almost certainly eliminated above 90%. This case shows how knowledge about an individual process in the treatment train can significantly impact the outcome. GAC data is sufficient enough to rule out a risk of high exposure to carbamazepine. At the same time, TFMSA may still pose a high risk. It is thereby necessary to either collect more information on the initial concentration or removal factors of previous treatments, or to add treatments to the exposure route that effectively remove TFMSA. A B C Figure 12: Removal factor distribution of drinking water treatment activated carbon for A: PFOS, B: TFMSA and C: carbamazepine D2.5Open source model for probabilistic human health risk assessment 35 The removal factor distribution of the combination of primary and secondary wastewater treatment (wwtt) is similar for all three substances, although 88 values from literature are available for PFOS and none for the other two substances. The distributions for the prior, the likelihood based on literature data, and the posterior of the PFOS removal factor during primary and secondary wastewater treatment is shown in Figure 13A. Although a lot of data is available, the posterior looks almost identical to the prior. This is due to the large range of observed removal factors in literature. Even the first 10 values of the database show a range between -105 % and 69% removal (note: negative values are treated as 0% removal in the HHEA model). Obviously, there is a huge variation in PFOS removal behaviour in different wastewater treatment plants (WWTP). Thus, literature data from one WWTP can rarely be used to predict the removal in a different WWTP, and consequently, the impact of the data-derived distribution on the prior is low. The posterior distributions of TFMSA and carbamazepine are very similar, although no removal data during primary and secondary wastewater treatment is available for either substance in the database. This example shows that high uncertainty about the transferability of data (highly variable literature data) is treated similarly to missing values by the HHEA model. A B Figure 13: Removal factor distributions of primary and secondary wastewater treatment for A: PFOS and B: TFMSA and carbamazepine. For the subsequent treatments in the exposure route, removal factors from literature indicated mainly poor PFOS removal. This leads to an overall higher risk of PFOS exposure compared to TFMSA. While the lack of knowledge is responsible for the categorization of TFMSA as a priority substance, the categorization of PFOS is based on a higher certainty of poor removal along the treatment chain. The HHEA model generally classifies both situations as risky. However, the high certainty of exposure reinforces this classification. 4.1.2. All substances: baseline scenario and scenario 1 The spider plots in Figure 14 and Figure 15 show the results of the baseline scenario and scenario 1. In the baseline scenario, all percentiles of final PFBA, PFPeA, PFHxA, PFBS, benzotriazole, and carbamazepine RQs are below 1, meaning no exceedance of the reference value is expected. For PFBA, PFPeA, PFHxA, and PFBS, the starting concentration is already below the reference value which explains the results (Table 6). On the contrary, the removal rates along the exposure route for benzotriazole and carbamazepine are sufficient enough to achieve acceptable concentrations in drinking water. D2.5Open source model for probabilistic human health risk assessment 36 The substances of highest concern and thereby the priority substances were PFOS, PFOA, TFMSA, PFPeS, PFDA, and PFNA, due to having an RQ > 10 at the 95th percentile, for which further scenarios according to Table 5 will be calculated. Figure 14: Spider plot of the water cycle baseline scenario (GAC as final drinking water treatment) The baseline scenario resembles scenario 1, in which only the final drinking water treatment was changed to ion exchange (IX). For most PFAS, this results in a slightly improved (i.e. lower) RQ. Nevertheless, the priority substances remain the same. Benzotriazole and carbamazepine are not as well removed as the PFAS. For both substances, exceeding their reference values cannot be excluded by using IX instead of GAC. D2.5Open source model for probabilistic human health risk assessment 37 Figure 15: Spider plot of the water cycle scenario 1 (IX as final drinking water treatment) 4.1.3. Priority substances: comparing different scenarios All scenarios for the six priority substances (PFOA, PFNA, PFDA, PFPeS, PFOS, TFMSA) were run. The scenarios have been given short, concise names for easier comparison. • Baseline: GAC, small clean river • Scenario 1: IX, small clean river • Scenario 2: GAC, large clean river • Scenario 3: GAC, large contaminated river • Scenario 4: GAC, small contaminated river The spider plots for each substance show a similar pattern (Figure 16). The use of IX instead of GAC results in a slightly more positive assessment for all priority substances. Not surprisingly, dilution with a large amount of clean surface water has the most positive effect on the result. However, even in this scenario, exceedance of the reference values cannot be excluded. What was unexpected was that for all pollutants considered here, whether the small river was contaminated or clean ("GAC, small clean river" vs. "GAC, small contaminated river") did not matter. There are two reasons for this: first, the starting concentration is set to the maximum found in the literature, which means that the specified contamination concentration (2 times the reference value) is hardly significant. Second, GAC is the critical treatment process for all priority substances. GAC is found at the end of the exposure route and therefore after the contamination event. D2.5Open source model for probabilistic human health risk assessment 38 A) PFOA B) PFNA C) PFDA D) PFPeS E) PFOS F) TFMSA Figure 16: Spider plots of all water cycle scenarios for the priority substances D2.5Open source model for probabilistic human health risk assessment 39 4.1.4. Priority substances: backwards estimation of tolerable starting concentration As mentioned in previous sections, there is no scenario for any substance in which compliance with the reference value is certain. One of the main reasons for this is the worst-case assumption for the initial raw wastewater concentration, which is either the maximum available value in the database or 10 times the reference value. The HHEA can also be used to estimate the maximum tolerable starting concentration for a given scenario and level of knowledge. The maximum tolerable starting concentration is the input concentration at which even the 99th percentile of the RQ distribution is < 1. To achieve this, the initial concentration is adjusted until all percentiles are green in the spider plot. For PFOA, this is the case at 15 ng/L (Figure 17). Figure 17: Spider plot of PFOA for different starting concentrations (Baseline scenario) The same procedure was carried out for the rest of the priority substances. The maximum tolerable starting concentrations are listed in Table 9. As more knowledge about the exposure route is gathered, this analysis can be repeated and the assessment becomes more realistic. Measured values can be directly compared to this benchmark value. This allows the HHEA to provide context for subsequent risk assessment and risk management even when there is extremely little knowledge about the substance and/or the treatment steps. A special case is the relative assessment of TFMSA. Instead of a tolerable starting concentration the simulation results in tolerable factor of a potential reference value. Thus, even substances without reference values (due to a lack of regulations or toxicological data) can be included and valuable information about the risk-based removal along the whole treatment train can be quantified and compared to other substances. D2.5Open source model for probabilistic human health risk assessment 40 Table 9: Maximum tolerable starting concentrations in the baseline scenario Substance Maximum tolerable starting concentration in ng/L PFOA 15 PFNA 1.4 PFDA 1.4 PFPeS 28 PFOS 10 TFMSA 3 Groundwater remediation The groundwater remediation exposure route shows to which extent contaminated groundwater can be treated to make it a safe resource for drinking water. Two different groundwater treatments, persulfate based chemical oxidation with Fe(VI) and ultrasonic cavitation, are compared in two different scenarios. The treated groundwater is then mixed with uncontaminated groundwater and is finally subject to the same drinking water treatment as in the semi-closed water cycle route (aeration, rapid filtration, GAC filtration). The starting concentration is based on on-site measurements of CS7 and CS6 in which groundwater was contaminated by AFFF application. For each substance the maximum value of available measurements is used as conservative starting concentration. As the exposure route is not linked to a specific site, the assumption of the dilution by groundwater is very simple and generic: treated groundwater is mixed with an equal amount of uncontaminated groundwater (average proportion of uncontaminated groundwater 50%, standard deviation = 10%). The exposure route can be seen again in Figure 18. Figure 18: Groundwater remediation exposure route. For both groundwater treatment processes, measurements for different PFAS were available. In case of the persulfate oxidation the pollutant concentration was measured after two separate injection events. The removal factors were supplied as average values (Table 10). For all substances, removal factors were derived after the first infection, while only some data was available after the second injection. For some of the substances removal factors between the two infections differed significantly, for others such as PFOS a high removal could be confirmed. . Both treatment trains were D2.5Open source model for probabilistic human health risk assessment 47 A spider plot of all compounds evaluated under the baseline scenario is shown in Figure 22. Among the 14 substances, four (PFDA, PFOS, PFOA and PFHxS) posed an unacceptable risk in most of the samples. Although within acceptable risk thresholds, PFNA would exceed 10% of the reference value with a 25% probability, while PFHxA would do so in 5% of cases. Figure 22: Spider plot for all 14 substances of water reuse route. Modelling results for all substances are in Annex I. See PROMISCES Deliverable D5.4 (Narain-Ford et al., 2024) for further information on DEP and PFBS exposure route and risk assessment. PFOA and PFOS were selected for further scenario assessment. Four scenarios were assessed, combining low/high starting concentrations (inlet) and poor/good performance of the electrooxidation treatment (E-peroxone), and results for PFOA are shown in Figure 23 and Figure 24. The assumed values for each scenario can be seen in Table 16. It should be noted that site values by (Meijide Fernández et al., 2025) show a wide deviation from the mean, even extending into negative ranges. This is mainly due to the fact that certain treatments apparently increase (rather than reduce) the concentration of certain compounds. This is because some compounds are generated during the treatment as a degradation process of other, more D2.5Open source model for probabilistic human health risk assessment 48 complex compounds, for which a decrease in concentrations and a positive removal factor are observed. Table 16: Changing parameters for the different water reuse scenarios. PFOA PFOS Comments Mean σ Mean σ Standard Raw wastewater concentration (ng/L) 26.7 75.2 61.7 78.9 All literature and site data considered. E-peroxone removal rate (wwel) (%) 28.4 41.3 14.2 60.3 All literature and site data considered. Scenario 1: Poor performance, High inlet concentrations Raw wastewater concentration (ng/L) 77.4 114.9 Value = 75% normal distribution. E-peroxone removal rate (wwel) (%) 0.6 -26.5 Value = 25% normal distribution. Scenario 2: Poor performance, Low inlet concentrations Raw wastewater concentration (ng/L) 1.0 8.5 Value = 75% normal distribution (minimum 1 ng/L) E-peroxone removal rate (wwel) (%) 0.6 -26.5 Value = 25% normal distribution. Scenario 3: Good performance, High inlet concentrations Raw wastewater concentration (ng/L) 77.4 114.9 Value = 75% normal distribution. E-peroxone removal rate (wwel) (%) 56.3 54.9 Value = 25% normal distribution. Scenario 4: Good performance, Low inlet concentrations Raw wastewater concentration (ng/L) 1.0 8.5 Value = 25% normal distribution (minimum 1 ng/L) E-peroxone removal rate (wwel) (%) 56.3 54.9 Value = 75% normal distribution. Note: Negative removal rates are considered to be zero by the tool. D2.5Open source model for probabilistic human health risk assessment 49 D2.5Open source model for probabilistic human health risk assessment 50 Figure 23: Modelled PFOA concentration profiles in the water reuse exposure route for the baseline scenario (Std) and4 adapted scenarios. The reference value is 6 ng/L PFOA. D2.5Open source model for probabilistic human health risk assessment 51 Figure 24: Modelled PFOA RQ for the water reuse exposure route for the baseline scenario (Std) and 4 adapted scenarios. Scenarios 1 and 3, which share high inlet concentrations, would result in unsafe concentrations under these conditions. Poor performance of the E-peroxone treatment (scenarios 2 and 4) poses a less critical risk to human health than probable high starting concentrations. The PFOS spider plot (see Annex I) shows a worsening in risk mainly due to a lower reference value, with no scenario showing RQ<1 above the 50th percentile. D2.5Open source model for probabilistic human health risk assessment 52 Nutrient recovery The nutrient recovery exposure route is based on pre-treatment of landfill leachate before it is mixed with municipal wastewater so that the final dewatered sludge from the wastewater treatment plant can be used as a fertilizer for agriculture. A total of 7 PFAS has been considered for this route: PFOA, PFHpA, PFHxS, PFHxA, PFPeA, PFBS and PFBA. No iPMT compounds were monitored in this route. PFOA was used to illustrate this route because it has been well studied in literature and data for it could be obtained. The exposure route can be seen again in Figure 25. Figure 25: Nutrient recovery exposure route. A more detailed definition of the exposure can be found in PROMISCES deliverable D4.5 (Lancioni et al., 2023) which described three different combinations of landfill and municipal wastewater treatment in Italy, and reported the sampling results from different campaigns. For the implementation as treatment train into the HHEA model, plant no. 3 from D4.5 was selected. In the plant, the landfill leachate is first pre-treated with flotation, sand filters, and ultrafiltration. The water is then treated with reverse osmosis (RO). The concentrate is further processed by flocculation with a sand filter and stripping before it is mixed again with the RO permeate. There are therefore two parallel routes within the treatment train, consisting of RO concentrate and RO permeate, which posed a challenge for the application of the HHEA model. It was decided to integrate the permeate (main flow: 1340 m³/d) into the exposure path and to add the concentrate flow (215 m³/d) via a mixing process afterwards. The effluent is then discharged into the sewer system and mixed with municipal wastewater before it reaches the municipal WWTP. Municipal wastewater treatment consists of primary and secondary treatment. It is assumed that PFAS that do not leave the wastewater treatment plant via the liquid phase and rather remain in the sludge (see section 4.5), which is then used agriculturally as fertiliser. As with the water reuse route, this is followed by two additional processes, dilution by irrigation water and concentration by evaporation, to reach the final concentration in the soil pore water. The concentration in the pore water is evaluated against the reference value for PFOA to obtain the reference quotient. The input variables are described in Table 17. Table 17: Variable definition for nutrient recovery route. Variable Value treatment_train ["wwt1", "wwnf", “wwro”, “dilpr”, “"dilww", "wwsl", "dilrw", "sepev"] substance “pfoa” case_study_name “nutrient_recovery” scenario_name “standard”, “cont_irrigation” input_matrix iww (industrial wastewater) n_runs 10 000 removal_factor_resolution 1000 (--> resolution of 0.1% removal factor steps) D2.5Open source model for probabilistic human health risk assessment 53 The initial concentration in landfill leachate was derived from on-site measurements (Lancioni et al., 2023) and literature values. The mean values implemented can be seen in Table 18. Table 18: implemented starting concentrations for the nutrient recovery route: Substance Number of data (N) Value (ng/L) PFHPA 9 9.8 PFPEA 4 9.8 PFBS 5 167.6 PFBA 6 4.8 PFOA 8 57.9 PFHXS 8 25.7 PFHXA 7 40.9 The first process (primary treatment) is based solely on literature values and leads to a reduction in the PFOA concentration below the calculated reference value in pore water. This is followed by mixing with the treated RO permeate, with literature and site-specific data from (Lancioni et al., 2023). Determining a fixed mixing ratio is challenging because it is site-specific and depends on the characteristics of both the RO permeate (removal efficiency) and the WWTP influent. The mixing conditions applied with the permeate (dilpr), as well as with municipal wastewater (dilww), to dewatered sludge (wwsl), with irrigation water (dilrw) and the concentration due to evaporation (sepev) in the soil are described in Table 19. In the table, x is the proportion of the added or removed water, c is the PFOA concentration of the external source of water in ng/L). RO concentrate data come from (Lancioni et al., 2023), the dilution by municipal wastewater is an estimate, and evaporation of irrigation water was assigned a conservative value (twice the value reported in the wetlands for the previous water reuse route) due to the lack of site-specific data. D2.5Open source model for probabilistic human health risk assessment 54 Table 19: Characteristics of mixing processes of the nutrient recovery route. Water xmean xsd cmean csd Comments Scenario: Standard RO concentrate 0.16 0.016 0 0 x standard deviation assumed to be 10% of mean value, process water characteristics based on ( Lancioni et al, 2023) Municipal wastewater 0.875 0.038 0 0 x entered as range between 0.8 and 0.95 Wastewater to dewatered sludge 0.30 0 0 0 assuming that all sludge contaminant migrates to 30% porosity Irrigation water 0.99993 5.5E06 0 0 x entered as range between 0.9993 and 0.9996 Evaporation water 0.21 0.015 0 0 x entered as range between 0.18 and 0.24 The dilution with irrigation water is >99.9% and is therefore extremely high. The calculation is based on the following assumptions: 1. The application of sludge as fertiliser is based on the P requirement of the soil. A high requirement of 60 kg P2O5/ha soil was assumed (worst case assumption). 2. It was also assumed that the P2O5 content of the sludge corresponds to 10%. This means that 6 kg/ha or 0.06 kg sludge/ m² soil are applied. 3. As the HHEA model works on a volume basis, the application amount was converted into a volume of 0.067 L /m² by using a sludge density of 0.9 kg/L. 4. 100 to 200 litres of water per m² are used within the lettuce growing period. This assumption leads to a volume-related proportion of irrigation water between 99.93% and 99.966%. Evaporation was also estimated as described for the water reuse exposure route. Concentration profiles show that PFOA is reduced to values below the reference value after the primary treatment (wwt1). Reference value for PFOA is set at 6 ng/L in pore water. Results can be seen in Figure 26. D2.5Open source model for probabilistic human health risk assessment 55 Figure 26: Modelled removal factors and concentration profile of nutrient recovery route for PFOA. A spider plot of all compounds evaluated under baseline (standard) conditions is shown in Figure 27. All 7 PFAS pose low risk in most of the cases (RQ ≤ 0.1 at ≥ 99th percentile). Modelling results for all substances are in Annex II. See PROMISCES Deliverable D5.4 (Narain-Ford et al., 2024) for further information regarding PFBS exposure route and risk assessment. D2.5Open source model for probabilistic human health risk assessment 56 Figure 27: Spider plot for all 7 PFAS in nutrient recovery route. Three substances (PFOA, PFHxS, PFHxA) were selected for further analysis in four additional scenarios. The scenarios combined poor/good performances of leachate treatment removal rate and nanofiltration. All scenarios considered poor WWTP performance. The assumed values for each scenario can be found in Table 20. The results thereof can be seen in Figure 28. D2.5Open source model for probabilistic human health risk assessment 63 5. Discussion and conclusions This deliverable describes the development and operation of the HHEA model. It describes the infrastructure, the operation and theory behind the probabilistic Bayesian updating approach, and explains how the model was applied to the 4 CE routes investigated in PROMISCES. For each CE route, the uncertainty of the risk of certain PFAS/iPM(T) of greatest concern were quantified. These compounds were determined by running a baseline scenario and observing which compounds had the highest RQ value. These compounds were subsequently analysed through 4 additional scenarios, to determine how changes in the exposure route would change the RQ. This was done so that the HHEA will improve understanding of decision making and risk management under uncertainty related to fate, exposure, and hazard. For the semi-closed water cycle exposure route, six priority substances which had RQs > 1 for at least one percentile in the baseline scenario were identified (PFOA, PFNA, PFDA, PFPeS, PFOS, TFMSA). By running scenarios with different sizes and background contamination of rivers and/or different final drinking water treatment technologies, dilution had the most significant effect on compound reduction. Additionally, it was revealed that GAC as final drinking water treatment could notably reduce compound concentrations regardless of whether the discharged effluent entered a small, clean river or a small, contaminated river. For the groundwater remediation exposure route, which was purely evaluated for modelling purposes, as contaminated groundwater would not be used for drinking water production, results were unlike the other exposure routes. The RQs of all evaluated compounds were multiple times greater than 1, due to the general assumption of a 5 µg/L concentration for each compound, as well as lack of knowledge of compound removal in the treatment processes. For this route, the ratio of the starting concentration to the reference value overwhelmingly dictated the RQ. Therefore, maximum tolerable starting concentrations which would lead to a maximum concentration of 5 ng/L in drinking water were calculated backwards. The more measurements there are confirming high removal (95%), the higher the maximum tolerable starting concentration could be: repeated measurements can have major impacts on the result. The maximum tolerable starting concentration was more sensitive to certainty of removal efficiency than to groundwater treatment. For the water reuse exposure route, six substances presented RQs > 1 for all percentiles (PFDA, PFNA, PFOS, PFOA, PFHxS and PFHxA). Subsequent scenarios analyses on PFOA and PFOS revealed that poor performance of the E-peroxone treatment posed less of a risk than high inlet concentrations for PFOA, whereas for PFOS, no scenario achieved RQ < 1 above the 50th percentile, primarily due to the lower reference value. For the nutrient recovery exposure route, all percentiles of all 7 PFAS compounds showed RQ ≤ 1. For subsequent scenarios analyses, especially high PFAS removal during membrane filtration resulted in similar low RQ values (RQ ≤ 1). The HHEA model addressed epistemic uncertainty, meaning uncertainty of the model via lack of training data, by differentiating how likelihood and posterior distributions are affected based on the literature data evaluated. One study presenting poor removal has a greater effect on the likelihood D2.5Open source model for probabilistic human health risk assessment 64 and posterior distribution than one study presenting good removal, thereby facilitating a risk-based approach where more evidence is needed to verify a high removal. To balance this out, when sitespecific data are included in the assessment, if the difference in removal factors is expected to average out over time, removal factors and input concentrations are combined randomly to prevent under or over estimation of exposure via worst case or best case scenario assessment. The influence of literature, measured, and modelled data on the HHEA model outcome is substantial, as both the availability and variability of data directly shape the posterior removal factor distributions and ultimately the risk classification. This was demonstrated using compounds PFOS, TFMSA and carbamazepine in the semi-closed water cycle exposure route. For PFOS, abundant but highly variable literature data results in wide, flat likelihood distributions, limiting the model’s ability to update priors and reducing predictive certainty across treatment stages. In contrast, the absence of literature data for TFMSA and carbamazepine is partially offset by measured data, which drives distinct model outcomes—low removal for TFMSA and high for carbamazepine—particularly in GAC treatment. This highlights how localized, high-quality data can significantly refine model predictions. Moreover, the model treats highly uncertain or variable literature data similarly to missing data, reinforcing the need for representative, transferable datasets. Ultimately, the model differentiates between risk driven by data gaps (e.g., TFMSA) and risk driven by well-documented poor removal (e.g., PFOS), demonstrating how both the quantity and quality of input data critically affect exposure assessment and prioritization. In the future, data from other models developed in PROMISCES (Groot et al., 2025; Zessner et al., 2025) providing concentrations or in silico parameters could also be integrated into the HHEA modelling. The Decision Support Framework (DSF) Tool developed in PROMISCES (https://promisces.eu/Results/Tools/PROMISCES+DSF.html) also links to the HHEA model. By providing exposure assessment as part of the DSF tool, users can access the pre-evaluated 4 exposure routes described in this report for an initial assessment of how PFAS are removed in the exposure routes. When users download the HHEA model to their local environment, all parameters can be changed, and additional data can be included in the HHEA assessment, providing users with a more site-specific assessment of PFAS and iPMT exposure. D2.5Open source model for probabilistic human health risk assessment 65 6. Literature Commission., E. 2020 DIRECTIVE (EU) 2020/2184 OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL of 16 December 2020 on the quality of water intended for human consumption. EPA., U.S. 1989 Risk Assessment Guidance for Superfund. Estado., B.O.d. 2003 Real Decreto 3/2023 por el que se establecen los criterios técnico-sanitarios de la calidad del agua de consumo, su control y suministro. Felizeter, S., Jürling, H., Kotthoff, M., De Voogt, P., McLachlan, M.S. 2020. Influence of soil on the uptake of perfluoroalkyl acids by lettuce: A comparison between a hydroponic study and a field study. Chemosphere 260. Groot, H., Sosnowska, A.W., P., Meesters, J., Wintersen, A., Zhiteneva, V., Devau, B., Valstar, J., del Val Alonso, L., Kittlaus, S., Liu, M., van Gils, J., Knoche, F., Markus, A., Oudega, T.J., Janssen, G., Sprenger, C., Baldwin, D. and Zessner, M. 2025 Deliverable D2.3 - Toolbox fate & transport modelling of PMTs in the environment. Health., M.D.o. 2013 Toxicological summary sheet for carbamazepine: CAS: 298-46-4. Lancioni, N., Blumenthal, E., Sgroi, M.F., F., Frugis, A., Lazzazzara, M.R., I., Valchev, D., Lubomirova, V. and Lincheva, S. 2023 Deliverable D4.5: Mass flow and fate analysis for PFAS from landfill leachate in Bulgaria and Italy. Martinez-Megias, C., Mentzel, S., Fuentes-Edfuf, Y., Moe, S.J. and Rico, A. 2023. Influence of climate change and pesticide use practices on the ecological risks of pesticides in a protected Mediterranean wetland: A Bayesian network approach. Sci Total Environ 878, 163018. Meijide Fernández, J., Herrero, J., Pérez-Estrada, L., Bosch, C., Pierina Orlando, D., Inés Bonansea, R., López de Alda, M., Llorca, M., Farré, M., Cano, A., Matamoros, V., Mas, T., Martínez, B., Pascual, J., Valero, S., Behnisch, P. and Besselink, H. 2025 Deliverable D4.3 - Recommendations on EAOP-CW based treatment system for water reuse applications. Ministerio de Medio Ambiente 2007 Guía Técnica de aplicación del RD 9/2005, de 14 de enero, por el que se establece la relación de actividades potencialmente contaminantes del suelo y los criterios y estándares para la declaración de suelos contaminados. Narain-Ford, D., Hof, M., Naus, F., Carter, K., Bosch, C., Sgroi, M., Boucard, P., Cavelan, A., Ribarova, I. and Hartmann, J. 2024 Deliverable D5.4 – Solution strategies to reach a non-toxic environment for 5 PM(T) uses from a system perspective and how they are perceived by the different stakeholders in the system RAIS. 2025 The Risk Assessment Information System., U.S. Department of Energy's Oak Ridge Operations Office (ORO). RIVM. 2023 Relatieve Potentie Factoren PFAS. Rückbeil, F., Sperlich, A., Dietrich, C., Hartmann, A., Kuckelkorn, J., von Wichert, A., Behnisch, P. and Besselink, H. 2025 Deliverable D4.1 - Performance and assessment of drinking water treatment trains for removal of PFAS and industrial chemicals. Schrenk, D., Bignami, M., Bodin, L., Chipman, J.K., del Mazo, J., Grasl-Kraupp, B., Hogstrand, C., Hoogenboom, L., Leblanc, J.C., Nebbia, C.S., Nielsen, E., Ntzani, E., Petersen, A., Sand, S., Vleminckx, C., Wallace, H., Barregård, L., Ceccatelli, S., Cravedi, J.P., Halldorsson, T.I., Haug, L.S., Johansson, N., Knutsen, H.K., Rose, M., Roudot, A.C., Van Loveren, H., Vollmer, G., Mackay, K., Riolo, F. and Schwerdtle, T. 2020. Risk to human health related to the presence of perfluoroalkyl substances in food. EFSA Journal 18(9). D2.5Open source model for probabilistic human health risk assessment 66 Silano V, B.B.J., Bolognesi C, Chesson A, Cocconcelli PS, Crebelli R, Gott DM, Grob K, Lampi E, Mortensen A, Rivière G, Steffensen I-L, Tlustos C, Van Loveren H, Vernis L, Zorn H, Cravedi J-P, Fortes C, Tavares Poças MF, Waalkens-Berendsen I, Wölfle D, Arcella D, Cascio C, Castoldi AF, Volk K, Castle L 2019. Scientific Opinion on the update of the risk assessment of di-butylphthalate (DBP), butyl-benzylphthalate (BBP), bis(2-ethylhexyl)phthalate (DEHP), di-isononylphthalate (DINP) and diisodecylphthalate (DIDP) for use in food contact materials. EFSA Journal 17(12). TCEQ. 2023 Toxicology of PFPeA. Umweltbundesamt. 2023 „Liste der nach GOW bewerteten Stoffe“. Uusitalo, L. 2007. Advantages and challenges of Bayesian networks in environmental modelling. Ecological Modelling 203(3-4), 312-318. Zessner, M., Baldwin, D., del Val Alonso, L., Derx, J., Devau, N., Janssen, G., Jou Claus, S., Kittlaus, S., Knoche, F., Liu, M., Markus, A., Valstar, J., Meesters, J., Meijers, E.M., Obeid, A.A.A., Oudega, T.J., Pathak, D., Sprenger, C., van Gils, J., Wicke, D., Wintersen, A., Zhiteneva, V. and Groot, H. 2025 Deliverable D2.4 - Guidance document on fate, transport and exposure for PMT's in the environment. D2.5Open source model for probabilistic human health risk assessment 67 Annexes Annex I. Water reuse simulation outputs Substance Baseline simulation PFDA D2.5Open source model for probabilistic human health risk assessment 68 Substance Baseline simulation PFNA D2.5Open source model for probabilistic human health risk assessment 69 Substance Baseline simulation PFHxS D2.5Open source model for probabilistic human health risk assessment 70 Substance Baseline simulation PFHpA D2.5Open source model for probabilistic human health risk assessment 71 Substance Baseline simulation PFHxA D2.5Open source model for probabilistic human health risk assessment 72 Substance Baseline simulation PFPeA D2.5Open source model for probabilistic human health risk assessment 79 Substance Baseline simulation PFOA D2.5Open source model for probabilistic human health risk assessment 80 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 81 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 82 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 83 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 84 Substance Baseline simulation PFOS D2.5Open source model for probabilistic human health risk assessment 85 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 86 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 87 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 88 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 95 Substance Scenario simulation D2.5Open source model for probabilistic human health risk assessment 96 Annex II. Nutrient recovery simulation outputs Substance Baseline simulation PFHpA D2.5Open source model for probabilistic human health risk assessment 97 Substance Baseline simulation PFPeA D2.5Open source model for probabilistic human health risk assessment 98 Substance Baseline simulation PFBS D2.5Open source model for probabilistic human health risk assessment 99 Substance Baseline simulation PFBA PFOA D2.5Open source model for probabilistic human health risk assessment 100 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 101 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 102 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 103 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 104 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 111 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 112 Substance Baseline simulation PFHxA D2.5Open source model for probabilistic human health risk assessment 113 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 114 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 115 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 116 Substance Baseline simulation D2.5Open source model for probabilistic human health risk assessment 117 Substance Baseline simulation