Introduction

Consumer products are a ubiquitous source of public exposure to a wide range of chemicals and as such have increasingly become the focus of research and regulatory risk assessment and management. Given the large number and variety of chemical substances present in consumer products and the variability in consumer use patterns, it is not possible to measure all potential exposures. It is therefore imperative that the exposure science community continue to rely on predictive tools to cover the likely range of exposures that may result from consumer product use.

This need has been met by the development of a variety of consumer product exposure models in research and regulatory communities. For example, in Europe, various models, such as the European Centre for Ecotoxicology and Toxicology of Chemicals Targeted Risk Assessment (TRA) tool, the European Solvents Industry Group Generic Exposure Scenario Risk and Exposure Tool (EGRET) model, and an updated web-based version of the National Institute for Public Health and the Environment in the Netherlands (RIVM) ConsExpo model [1,2,3,4], were developed and are used in registering chemicals under the European Union’s Registration, Evaluation, and Authorisation of Chemicals (REACH) regulation. In the US, the Environmental Protection Agency’s (EPA) Office of Pollution Prevention and Toxics recently upgraded the Consumer Exposure Model (CEM) by including 15 additional models to cover various exposure and use scenarios from both product and article uses for their regulatory risk assessments [5]. Also, EPA’s Office of Research and Development’s National Exposure Research Laboratory has developed the Stochastic Human Exposure and Dose Simulation model (SHEDS-HT) from its full SHEDS--multimedia and multipathway to accommodate high-throughput assessment of exposure from consumer products for chemical prioritization [6, 7].

These models have different objectives (screening vs. in-depth assessment), capabilities (single chemical vs. multiple chemicals), product representation (categories of products vs. single product/use), and information requirements (number and type of input parameters, including use of point values or distributions). Some models address formulated products as well as articles; others focus just on formulated products. Screening level models such as the TRA are designed to be inherently conservative (intended to overpredict rather than underpredict exposures); whereas higher tier models such as ConsExpo allow for more accurate and even probabilistic exposure predictions. The SHEDS-HT model assesses population exposure, which includes product users and nonusers, whereas the other models focus on exposure to the consumer product user. As available models are designed for different applications and different tiers of exposure assessment (i.e., screening to higher tiers), key model components vary including: platforms on which they operate, approaches to representing products, algorithms for estimating exposure, deterministic versus probabilistic estimates and exposure factors, ability to modify defaults, the calculated exposure metrics, periodicity and/or duration of exposure, etc.

Because of these differences, when the models are used in exposure and risk assessment of chemicals for consumer products, they can provide disparate results and the origin of these differences may not be obvious. The amount of effort to understand the causal factors and their relative influence on the results may be beyond the time constraints and in some cases the expertise of the model users. These differences can also contribute to inter-user variability which can cause additional disparity in results and lead to lack of confidence in the models. In a tiered application of exposure models, it is particularly important to understand if these differences represent a refinement of the exposure estimate, reduction in the uncertainty in the exposure scenario, or arise from other factors.

This Perspectives Paper highlights areas where the authors believe, with reasonable effort and focus, the exposure science community could promote consistent practices that would facilitate more systematic evaluation and a wider range of applicability of consumer product exposure models and their predictions. The insights identified are illustrated by recent comparison of predictions for four consumer products using five consumer models (see Supplementary Information) as well as the authors’ collective experience in exposure assessment. These efforts could increase the applicability of existing models for other product use situations and geographies by enabling assessors to transparently understand the relative contribution of model algorithms and exposure factors to the exposure. For example, if the assessor determined that the scientific basis of the model algorithms applies to their consumer product use scenario then updating the exposure factors to represent the population of interest would be appropriate. Also, this effort could focus opportunities to use emerging and recently developed technologies for collecting exposure information that would significantly improve understanding of, and confidence in, consumer exposure models and predictions [8]. For example, if algorithms, parameters, or exposure factors that are key contributors to model predictions and differences across models could be identified, then emerging technologies, such as personal sensors or personal collection devices of exposure factors and microenvironmental factors, could be targeted to improve the understanding of these key contributors. It is therefore very timely to consider the insights gained by the authors and use these to identify the optimal ways to achieve this understanding and confidence.

In brief, the insights identified and discussed include:

  1. (A)

    Consistency among models in describing types of consumer products, exposure routes, and associated exposure scenarios.

  2. (B)

    Consistent definition of model exposure metrics.

  3. (C)

    One model vs. application of multiple models within a tiered approach.

  4. (D)

    Separating model algorithms from databases of exposure factors information.

  5. (E)

    Evaluating model predictions vs. measured exposure data.

Indeed some of these observations and recommendations are not new but reinforce those put forth previously by the European Commission sponsored harmonization of consumer exposure models on a Global Scale project [9]. However, we believe that our experience builds on and expands on these recommendations in a very meaningful way.

The variability and uncertainty in predicted consumer exposure reflects the diversity of product uses, scenario descriptions, and exposure factors amongst models. For example, consumer product uses can be highly variable across different age groups. Furthermore, insufficient knowledge of the underlying mechanisms influencing potential exposure and how to represent these in the model’s algorithms can lead to uncertainty in consumer product exposure estimates. The exposure community should consider how to implement practical ways to address this diversity while retaining features such as flexibility and ease of use. There are several papers published that discuss the topic of variability and uncertainty in exposure models [4, 10,11,12]. It is beyond the scope of this Perspectives Paper to address these topics in detail. The focus of the following sections is to offer more specific discussion and recommendations on the five key insights identified above.

Discussion

The importance of consistency among models in describing types of consumer products and associated exposure scenarios

A key learning from the model comparison study (Supplementary Information, Table 1) and other experiences is that it is not straightforward to match the product types chosen for the comparison with the products available in the models. For example, in the model comparison, for some of the products the matching was fairly straightforward (e.g., all-purpose liquid cleaner). However, for others there was no reasonable product match. Instant action air-freshener was used in the product descriptions of TRA, EGRET, and CEM but for SHEDS-HT since there was no product with this name, air-freshener spray was chosen after considering the use pattern. ConsExpo does not have a default scenario for instant action air-fresheners. For this product type, professional judgment was used to identify the product type which would result in closest agreement to the exposure pattern associated with instant action air-fresheners. In this specific case, this was the air care products, pest control product with electrical evaporators. This step of choosing the appropriate matching product in cases when there was no obvious match was time consuming and required detailed understanding of the products as represented in the models and their associated product use patterns, especially since among the models there is not a common naming for products that appear to be similar (e.g., instant action air-fresheners). This detailed level of information might not be available to all assessors.

Table 1 Product scenarios chosen in models to represent product types.
Full size table

Even when the product types had a match that seemed appropriate, the exposure routes and scenarios for the chosen product may differ among the models (Table 2). SHEDS-HT was the only model that included indirect ingestion of chemicals route from hand-to-mouth transfer of chemicals left on surfaces, which in all cases was negligible (see Supplementary Information). Also, CEM did not include dermal exposure by default for all-purpose spray cleaner but the other models did. The lack of dermal exposure in CEM included other spray products, e.g., spray fixative and spray coatings. The route of exposure for instant action air-freshener in all the models was inhalation only, except for SHEDS-HT which included dermal exposure. Even when the exposure routes appear to be the same among the models, there are differences in the actual scenarios for which the exposure is estimated. For example, in ConsExpo for spray fixative (do it yourself products/glues/spray glue—application), the scenario describes gluing a poster to a wall or into a frame with a default amount used of 255 gm/use [13]. The CEM scenario (spray fixative and finishing spray coatings) however, describes an art project with a default amount used of 10–40 gm/use [14]. Furthermore, the number of uses (12/year for ConsExpo and 7/year for CEM) and the exposure duration (240 min for ConsExpo and 15–20 min for CEM) differed among the exposure models. Therefore, it is often complicated to compare the resulting exposure predictions. In order to remove (or separate out) the impact of daily frequency of use in the model comparison, in the model comparison study (Supplementary Information) this parameter was set to 1/day. This allowed us to explore the sensitivity of the model’s estimates of exposure to other factors. However, for a regulatory assessment, this choice of 1/day may not be appropriate.

Table 2 Exposure routes considered for each product type.
Full size table

Obviously, these differences in product type, exposure routes, and exposure scenarios can result in different estimates of exposure. If these are not explicitly recognized by the assessor and accounted for in comparing exposure results, then conclusions on the level of exposure could be in error.

Recent advances, such as the development of the OECD internationally harmonized functional, product and article use categories [15], when implemented, will help to address some of these difficulties. Use of harmonized use codes to cross-reference consumer products within the models would improve the consistency and efficiency in matching product types among the models. Furthermore, use of a harmonized use code system would make it easier to communicate what products had been assessed and readily identify if a consumer product of current interest has an existing assessment. This would greatly improve the ability to communicate which products have been assessed among assessors globally. It is not recommended to limit the products available within any model—including those used in regulatory assessment—to these harmonized product use codes. Rather the goal would be to provide aligned product and use descriptions as options available to exposure assessors, starting with these use codes.

A telescoping categorization scheme, such as is used in commercial food definitions (Food Codes) may be a useful approach [16]. The telescoping approach would start with a small list of general consumer product categories. The next tier would refine this first tier list by including specific product forms and uses with the goal of refining consumer exposure estimates. Each additional tier, if needed, would represent an additional level of refinement and detail regarding product forms and uses that impact consumer exposure. This type of approach can be seen in the REACH product codes and subcategories already built into tools such as TRA. In the TRA additional scenarios can be added into the base set of available ones. Recently, EPA developed hierarchical product use categories (PUCs) by mapping product use and exposure scenarios (including relevant exposure routes) to the consumer products in EPA’s chemicals and products database [17]. An example of this type of hierarchical scheme would be a general category of household cleaners refined to hard surface cleaners subcategory which is then further refined to bathroom surface cleaners. The exposure estimates from models which have been parameterized using these PUCs can then be easily compared or transparently communicated without extensive work delving into the details of the models.

Similarly, the addition of an agreed-to list of the relevant exposure routes for uses and scenarios for each of the product use codes would increase the transparency and comparability of these exposure assessments globally. The Specific Consumer Exposure Determinant (SCED) templates developed and applied for REACH can serve as an illustration. These or other templates could be expanded to provide documentation for which routes were chosen or considered negligible. Following this type of consistent, transparent format will make it easier for assessors to determine if the routes selected are appropriate for the scenario they are interested in. This could also help assessors in different geographies “align” assessments more confidently. Developing a harmonized list of exposure routes and scenarios would also facilitate the collection and communication of the minimum and common data elements needed to run lower tier models such as TRA. The SCED templates [18] have the advantage of being aligned with the input parameters needed to run the REACH consumer exposure model. They have a base set of fields which match TRA input values and include the possibility to add additional fields required by ConsExpo. They could potentially be augmented with optional information elements needed for higher tier model analysis.

Harmonized product types (through use codes), relevant exposure routes, and scenarios would provide the foundation for discussion and evaluation of how regional differences in consumer habits and practices could impact exposure, thereby enabling application of existing assessments to other regions.

Consistent definition of exposure metrics

Ultimately the goal of the consumer product exposure models is to provide exposure metrics that can be compared with hazard benchmarks to determine the potential for risk. However, exposure metrics differ across the models (Table 3). Even in some cases where they are identically named, they do not necessarily represent the same exposure metric. For example, ConsExpo and CEM both provide estimates of peak exposure. In CEM, this metric represents the highest instantaneous exposure (calculated in 10-s increments); in ConsExpo it represents the highest 15-min exposure. If the assessor is not aware of this difference in definition, a simple comparison of Peak Exposures could lead to conclusions that are incorrect regarding relative exposures. This could in turn potentially result in different conclusions of risk.

Table 3 Exposure metrics calculated by models.
Full size table

Secondly, the exposure metric chosen and the units of exposure used can result in differences in the rank order of model predictions among the models and ultimately the subsequent exposure and risk assessment. The choice of metric and the necessary modifications to the models, e.g., forcing a use of once per day, could impact findings. For example, Zaleski et al. [2] found that a different rank order may be seen if the inhalation event concentrations in mg/m3 (the metric used for REACH) are used instead of single event daily exposure in mg/kg/day presented here. In addition, there is an impact on the exposure estimates of the frequency-of-use defaults in the models, which can include multiple uses per day. For example, in the EGRET tool the default number of uses for the instant action air-freshener is four times per day; therefore, total daily exposure in mg/kg/day would be four times the single product use exposure mg/kg/day. This would result in a different order of model predictions on a daily total exposure basis, which includes number of uses, than on the event total daily exposure basis. Thus, to facilitate direct comparison of the exposure metrics between models and to ensure appropriate use of the exposure metrics in risk assessment, a common set and definition of these metrics is needed. A basic set of metrics with a description would allow consistent application of these exposure metrics in assessments. For example, “peak exposure” metric could be more explicitly described by including the averaging time in the metric description, e.g., 15 min peak exposure. For example, this approach is consistent with occupational hygiene practice where for example, an 8- or 24-h time weighted average would be specified with the exposure concentration. This list could be continually expanded on as need arises.

Furthermore, it would be useful for models to provide results in multiple metrics that are commonly used in a safety or risk assessment. For example, both the dermal event exposure metric (mg/cm2) and the daily exposure metric (mg/kg/day) are useful for evaluating potential local and systemic health effects but not all of the models provided these exposure metrics (Table 3). Also a variety of exposure metrics will facilitate the evaluation of risk consistently without the assessor having to perform additional extrapolations. For example, the ConsExpo inhalation results in basic mode, without specifying an absorption model or changing the default to one use per day, estimates a metric of event exposures (mg/m3). The day of use exposure for that single event (mg/kg/day) was calculated using default body weight and inhalation rate as described in Supplementary Information.

Therefore, a further recommendation would be that all exposure models provide a common set of exposure metrics as part of their output. Like product code harmonization, harmonization of exposure metrics would ideally be an international effort with consideration of the needs among the various regulatory bodies and the relevant toxicity endpoints. This could initially include consistent definitions and information on how to convert among the exposure metrics that are currently provided by the models. It is not suggested to confine reported exposure estimates to a limited set of harmonized outputs, but rather to ensure that this minimum harmonized set of metrics be included - along with any others appropriate for that model. Ultimately the assessor would need to select the appropriate metric for use in their decision-making context. As a community, it is important that we use terms that are clearly defined and specific enough that they mean the same thing to everyone including those outside the exposure science community. The meanings of these metrics should be unambiguous to risk assessors and scientists so that they are employed appropriately. This is particularly important when these exposure metrics are combined with hazard information in developing risk estimates that could lead to risk management decisions.

Multiple models vs. one harmonized model

Given the previous discussion, one could conclude that there is a need for one agreed-to model for consumer product assessment. However, work with consumer product models by these authors has made it evident that use of multiple, fit-for-purpose consumer exposure models generally provides more value to assessors. Also, models that fall along multiple tiers from screening level tier to higher tier enable exposure predictions to align with data or resource availably while being tailored to the requirements of the assessment.

There are different approaches to representing the underlying science governing chemical movement in the environment and into/within receptors; therefore, models may differ in algorithms and input values because different valid approaches are chosen in the models. For example, dermal exposure is represented in all the models by a similar algorithm that is generally agreed to among assessors. Therefore, the observed differences in exposure (see Figs. A1-I–1-IV in Supplementary Information) are easily associated with differences in the amount of chemical loaded on the skin, use frequency and surface area exposed as shown in previous work by these authors. If a more refined estimate is needed that considers factors such as dermal permeability or the effect of protective devices on exposure then other models could be used. IH SkinPerm [19] is an example of a model used widely in the industrial hygiene community to estimate dermal exposure. This particular model is also available in four different languages and as such is a good demonstration of how one model can have wide geographic applicability.

However, the science and the appropriate algorithms for estimating inhalation exposure is still in evolution; therefore, the models considered in the model comparison had different algorithms and included different factors. The ConsExpo and CEM models, for example, used different algorithms to describe the steady or time-varying emission rates to the air volume and the countervailing removal rates of ventilation and non-ventilatory mechanisms such as aerosol sedimentation or adsorption to surfaces. Specifically, the models are quite different in their basic approach to modeling the airborne VOCs during spraying and the airborne particulate from spraying. With these differences come differences in specific model predictions for products with similar use patterns and the exact origin of these differences is much more complex to understand and evaluate. Therefore, using the different models, as illustrated in the Supplementary Information, provides a more robust understanding of the likely range of inhalation exposure.

The diversity of approaches also reflects both scenario and model uncertainties that challenge exposure characterization. For example, if the model relies on measured data to characterize product use, does that data apply to the new exposure scenario being evaluated? If the model relies on information generated in a given geography, does the model adequately represent the population in the exposure scenario, especially if that population resides in a different geographic region?

It is also evident that several improvements could enable broader application of existing assessments to new situations, which could in turn streamline regulatory processes, enabling more substances to be assessed in a shorter timeframe. Most importantly, there needs to be improved and consistent descriptions of the valid range of application of the different exposure models and algorithms which include: (1) the evaluation of the fitness of the model for the purpose of the assessment being conducted, (2) the ability/confidence to make predictions with limited data, (3) the limitations of these predictions, and (4) when warranted, when and how any screening level predictions can be refined. Even limited comparisons of model predictions to measured data may be able to provide insight into the degree of confidence that can be placed on the predicted values.

In some cases model documentation includes a good description of the domain of applicability (i.e., range of vapor pressures that can be assessed in the model) but even when this is the case, sometimes this information may be difficult to locate. There are many ways that this information could be highlighted, such as providing a link to the documentation in the model or output, or having summary information in a help screen available within the model, amongst others. Building in the ability to conduct sensitivity analysis would be an effective way to help exposure assessors identify where screening level predictions could be refined, by identifying those input parameters with the largest impact on the results. These descriptions would allow differences in exposure potential due to different use patterns to be identified and the assessor to judge if the selected model is appropriate for the population of interest. Finally, this would also allow resources to be focused toward developing and archiving model input data, which would ultimately result in more representative exposure estimates.

Whereas multiple models can give additional information, where there is a common purpose for use moving toward a common core set of consumer product exposure models may be appropriate. With a common model, exposure datasets could become more harmonized and potentially applied more widely. The value of harmonized datasets and tools is especially important given the increased capabilities in data analytics as well as the easier and less costly acquisition of exposure data. For example, for some screening level assessments, the TRA and ConsExpo give the same exposure estimates at lower levels. Higher tier assessments can be conducted by simply modifying or changing parameters. Progress in this area could have significant positive impact on the future of consumer exposure and risk assessments globally.

Predictive exposure models need to cover the full range in consumer contact scenarios with substances during and subsequent to product use. Implementing these recommendations could lead to: (1) development of common datasets, (2) development of suites of modeling tools with wider international application, (3) collecting more data representative of local use scenarios, (4) more consistent interpretation/application of results, and (5) facilitation of future data collection and expanded use of these data. In addition, this would assist meta-analyses or linkage between datasets both now and in the future.

Separating model algorithms from databases of exposure information

Model algorithms and input databases are sometimes directly connected and non-editable (ex., built-in data), so that it can be challenging or impossible to independently evaluate their impact on the exposure metric. There are multiple advantages to built-in databases which include the necessary information to run the model. The primary advantage is the time saving for all users, who would otherwise need to identify and enter this information. It also facilitates consistency across users, when assessments use a common dataset. However, transparency of built-in data is an essential model feature, so that the model user can evaluate the model fitness for their specific application.

It was not always readily apparent if differences in results were due to differences in the model algorithms or to differences in exposure factors and other model parameters. For a given model, a complete listing of the input data and associated algorithms may not be available. Unpublished experience includes an evaluation of the methods used within ConsExpo and CEM to estimate exposure to airborne VOCs and particulates during spraying. Determining the relative contribution of differences in the algorithms versus the underlying exposure factors and model parameters to the exposure estimates from these two models was challenging and was not completely able to be evaluated. This interconnectedness of data and algorithm is a challenge even for an individual model and further compounds the difficulty in comparing models. Ultimately, separating algorithms and datasets would address this challenge. Because consumer habits and practices change over time and vary by region, the ability to separate the input databases from the model algorithms would facilitate greater geographic application of a model when algorithms are appropriate but the input data would need to be modified to represent a new target population or temporal situation.

Model transparency sufficient to support model evaluation should be a general goal. When probabilistic analyses are done, directly matching model output with the input variables is further complicated. For example, in the Supplementary Information, product-specific individual intake fractions, which requires an estimate of both total volume used per event and total individual exposure per event, were compared. However, it was not possible to match the 95th percentile SHEDS-HT intake fraction with the exact input value because product use amount is specified as a lognormal distribution with a mean value and coefficient of variation. For this particular analysis, it would have been useful if the model output included a file of all the input values associate with a given exposure prediction. As exposure science continues to progress, and the use of multivariable probabilistic approaches increases, it will be important to consider what level of direct linkage between model input and output is appropriate to document. It would benefit the user to have access to a straightforward reporting of parameter value(s) used and identification of the algorithm and/or equations used by each model to generate the exposure estimate, at least as an option. We recognize that this level of detail may be cumbersome and not desirable for all users.

Comparison of measured exposures with model predictions

Ultimately, the greatest improvement in the confidence in consumer exposure model predictions will be achieved by comparing analytical measures of exposure to model predictions. For example, the wall-paint algorithm in EPAʼs CEM model was tested vs. empirical data [5]. In addition, the REACH models were evaluated using the USEPA wall-paint studies [20] and for one cleaning agent [21] where sufficient information on both air concentrations and exposure conditions was found. This type of comparison can help define the model application domain for the types of chemicals and use scenarios. However, this comparison may not be straightforward for the following reasons:

  1. (1)

    There are uncertainty and variability in monitoring data. Monitoring data usually reflects the conditions under which it is collected and these conditions may or may not represent the full range of potential consumer product exposures.

  2. (2)

    To be useful for model assessment, measured exposure data as well as detailed information on product use and consumer contact with the product are needed, so that the model inputs can be aligned with the conditions of the exposure measurement.

  3. (3)

    Monitoring data and studies are limited because they are usually difficult and expensive to acquire.

With new technology, a large amount of exposure data from various exposure scenarios may be collected or simulated quickly. Recently, to expand the ability to collect data to evaluate consumer exposure models, a robot-painting study was completed as a proof of concept; measured air concentrations were consistent with model predictions, with lower tier models being more conservative and measured air concentrations similar to higher tier models predictions parameterized to the painting conditions [8]. Overall, appropriately designed studies which measure the exposure from representative product use scenarios, along with the collection of associated contextual information needed to run models, can be used to evaluate and improve the exposure model predictability.

Biomonitoring data provides useful information on individual exposures, but these values represent an individual’s total exposure via all sources, typically obtained without detailed contextual information on the sources of exposure to the individual. Comparison of biomonitoring data with predicted external exposures also requires conversion of internal levels to associated external exposures. A study recently published by Aylward, et al. [22] demonstrates the value of collecting detailed diary information on consumer product use along with biomonitoring measurements. These data enabled measured exposures to be compared with predictions of multiple models, including several in this comparison. The results showed that 90% of the model predictions for most of the models were within a factor of 10 of the observed exposures and ~30–40% of the predictions were within a factor of 3. At a population level, biomonitoring data has also been used to benchmark performance of high-throughput exposure models that include consumer exposure pathways (such as SHEDS-HT) by reconstructing exposures through reverse dosimetry techniques [23,24,25]. In Europe, the HBM4EU program outlines methods for reconstructing exposures from biomonitoring data and provides recommendations for combining measured and predicted exposure data [26].

The application of current and emerging technologies which can facilitate co-collection of data on exposure sources along with biomonitoring data could provide reductions in uncertainties and improvements in model performance. Ideally, consistent database formats and structures established within the exposure science community, at the initiation of such data collection, would maximize data sharing. For example, approaches such as the globally harmonized formats of the multinational time use study [27, 28], in combination with data elements needed as model input, could provide a useful starting point. Anticipating future demands for better incorporation of in vitro, in vivo, and ambient environment data, approaches more intimately linking exposure and hazard data at common nodes (ex., Aggregate Exposure Potential and Adverse Outcome Pathway frameworks [29]), may also enable modular database design promoting interactions within a risk assessment framework.

In addition, further comparisons of predicted to measured data will help understand if, and to what extent, the uncertainty and the variability of exposures are captured by different models. For example, the intent of using a screening tool is often to exclude low-risk scenarios from further evaluation. Therefore, the results of a screening exposure assessment would generally be desired (and expected) to exceed the expected (or a measured) value in order to provide confidence that the scenario being described is unlikely to present a risk. Thus, while the confidence in the actual predicted value is low (low accuracy) there would be confidence that the predicted exposure exceeds the actual exposure and thus it is appropriate for using for screening purposes. Higher tier tools would be expected to characterize better the range of exposures (i.e., precision) as well as have improved accuracy (i.e., predicted X%ile of exposure is closer to measured X%ile of exposure).

Conclusion

The model comparison activity described in Supplementary Information combined with published and unpublished information highlights several key areas that could improve the use of and confidence in consumer exposure models:

  1. (1)

    Consider the adoption of harmonized product use codes that can be used independent of or in conjunction with model-specific exposure scenarios. Harmonized use information would allow for comparison of “standardized” exposure scenarios across models, help reduce duplication, and expand global use.

  2. (2)

    Consider the development of a “standard” list of exposure metrics that could be used independent of or in conjunction with model-specific reporting metrics. It would be important that these metrics be consistently defined and documented for ease of comparison across models. For example: what does peak exposure mean; perhaps it should be described in combination with averaging time, e.g., 15-min peak exposure.

  3. (3)

    Where feasible, separate out databases of parameter values i.e., exposure factors, from the algorithms and programming that use them. This would allow for greater flexibility in conducting sensitivity analysis and updating parameters when new information becomes available. The utility of having preset default selections is noted and having existing tables of parameter values (i.e., that can be selected in a dropdown or uploaded from tables) can streamline the assessment process and reduce data-entry error. However, models should still ideally allow for custom values to be used as well.

  4. (4)

    Continue to generate measured data against which to corroborate predicted exposures. Exposure measurements need to be collected with sufficient contextual information so that the model(s) can be run to evaluate the consistency between model predictions and the exposure measurements. This will help characterize both uncertainty and variability in the exposure estimates. It is also important that where measured data underpin semi-empirical models that it be possible to update or augment that underlying information. Combining detailed surveys that include product use information, time and activity patterns, and food ingestion diaries would be useful for this purpose.

This project has demonstrated the value of side-by-side comparison of exposure predictions from different models that are used in regulatory chemical risk assessments for formulated consumer products. Future work focusing on wider comparison of products and ideally with comparison to monitoring data would add to this knowledge base. But ultimately the authors believe that addressing the ideas shared above will result in the greatest improvement in the confidence of model predictions and accurate communication of consumer product exposure assessments globally. These observations and recommendations are consistent with and build on the findings of a previous project evaluating the potential for harmonization across consumer models [9].