Quantitative microbial risk assessment of Greywater on-site reuse

Recycle domestic greywater for on-site non-potable uses can lessen the demand on potable water and the burden on wastewater treatment plants. However, lack of studies to assess health risk associated with such practices has hindered their popularity. A Quantitative Microbial Risk Assessment was conducted to estimate the public health risks for two greywater reuse scenarios: toilet flushing and food-crop irrigation. Household greywater quality from three sources (bathroom, laundry and kitchen) was analyzed. Mathematical exposure rates of different scenarios were established based on human behavior using Monte-Carlo simulation. The results showed that, greywater from all three household sources could be safely used for toilet flushing after a simple treatment of microfiltration. The median range of annual infection risk was 8.8 × 10−15–8.3 × 10−11 per-person-per-year (pppy); and the median range of disease burden was 7.6 × 10−19–7.3 × 10−15 disability-adjusted life years (DALYs) pppy. In food-crop irrigation scenario, the annual infection risks and disease burdens of treated greywater from bathroom and laundry (2.8 × 10−8, 4.9 × 10−8 pppy; 2.3 × 10−12–4.2 × 10−12 DALYs pppy) were within the acceptable levels of U.S. EPA annual infection risk (≤10−4 pppy) and WHO disease burden (≤10−6 DALYs pppy) benchmarks, while kitchen greywater was not suitable for food-crop irrigation (4.9 × 10−6 pppy; 4.3 × 10−10 DALYs pppy) based on these benchmarks. The model uncertainties were discussed, which suggests that a more accurate risk estimation requires improvements on data collection and model refinement.


Introduction
In pace with the human population explosion, water shortage is regarded as one of the most concerned issues around the world. About 1 billion people worldwide now don't have access to safe drinking water (Connor, 2015). Development of new water resources is imperative. Instead of seawater desalination or drilling deeper into groundwater aquifers, collecting and recycling municipal wastewater for on-site non-potable uses can curtail demand of freshwater as well as lessen the burden of wastewater treatment plants. Such approach is likely to be a feasible and eco-friendly direction.
Municipal wastewater is generally divided into yellow water, brown water and greywater. Among them, yellow and brown water refer to urine and fecal sewage. Greywater, including streams from showers/ baths, wash basins, laundry, kitchen sinks and dishwashers, is generally defined as urban wastewater without any pollution from toilets (De Gisi et al., 2016). Of these three, greywater is the most suitable for water reuse because of its large volume and low concentration of pollutants (DeOreo et al., 2016). The invention of source separation system makes it possible to reuse domestic greywater on-site as a new water supply (Larsen and Gujer, 1997). It segregates wastewater streams at their sources and treats them separately according to their qualities, so that water reuse can be more economical and less complicated.
Toilet flushing is one of the most frequently discussed application of greywater reuse, through which household indoor water demand can decrease by N20% (De Gisi et al., 2016;DeOreo et al., 2016;ENEA, 2002). Another common route of greywater reuse is outdoor irrigation (Lubbe et al., 2016;Pandey et al., 2014). Across many regions in the United States, the amount of outdoor water consumption accounts for N50% of total water use in a single household (DeOreo et al., 2016). Previous study indicated that greywater from bathroom and laundry not only can meet demands of toilet flushing entirely, but also can satisfy part of irrigation demand if necessary (National Academies of Sciences and Medicine, 2016).
Constituents in greywater are related to diverse factors, such as source of water supply, household activities and water-consuming installations (De Gisi et al., 2016). Generally, domestic greywater is generated from bathroom, laundry and kitchen; and their qualities can differ significantly (Li et al., 2009;Oron et al., 2014). Previous studies of greywater quality showed that bathroom greywater contains a low concentration of biochemical oxygen demand (BOD), nitrogen and phosphorous, and is regarded as the cleanest stream of greywater (Bodnar et al., 2014;Li et al., 2009;Oron et al., 2014). Due to the addition of laundry detergents, laundry greywater has an elevated alkalinity, pH and high loads of sodium, nitrogen, phosphorous and surfactants, but its level of BOD is relatively low. In contrast, kitchen greywater, which comprises oils, fats and food debris, shows high concentrations of BOD, nitrogen, phosphorous as well as turbidity, and is considered as the most polluted greywater. Despite the absence of urine or fecal contaminations, all streams of greywater, however, contain microbial contaminants. Even in bathroom greywater, the concentration of total coliforms can exceed 1 × 10 7 CFU/100 ml (De Gisi et al., 2016), which poses a potential health threat during its practical reuse application.
Diverse reuse routes can result in different exposure scenarios of greywater to human bodies, which ultimately determine the magnitude of health risk. Toilet flushing and outdoor irrigation represent two distinct routes of greywater exposure: respiratory tract vs. digestive tract. For toilet flushing, greywater aerosols are produced after a single toilet flushing. The aerosols can be inhaled into human respiratory system, which brings in harmful constituents simultaneously (Lim et al., 2015). As for irrigation, it's been estimated that 31% of U.S. households participated in food-crop irrigation (Butterfield, 2009). The intake of raw home-grown produce provides a direct pathway for greywater retained on crops to enter human bodies (Lim and Jiang, 2013). Although exposure to hazards doesn't necessarily equal an unacceptable threat, such facts do indicate the existence of potential risks.
Previous studies claimed that untreated (or inadequately-treated) greywater is not recommended for either potable or non-potable uses (Etchepare and van der Hoek, 2015;Kuru and Luettgen, 2012;Maimon et al., 2010). Proper treatments are necessary to remove pollutants in greywater, especially microbes. Existing greywater treatments include physical/chemical, biological and ecological processes. Physical/chemical processes are proficient at removing suspended solids but cannot guarantee an adequate reduction of organics and nutrients (Brewer et al., 2001;Gerba et al., 1995;March et al., 2004). The installations of these technologies are relatively inexpensive and easy to operate. Biological processes can ensure satisfactory and stable effluent quality but are relatively complex to operate and expensive to set up (Abdel-Kader, 2013;Liu et al., 2005). Ecological processes, although being the most economical and environment-friendly technologies, commonly require large areas and long storage time that are often not met in urban settings. Comparatively, it is generally agreed that physical/chemical processes are best suited for on-site greywater reuse in most conditions with fair expenses. However, the microbial safety of using greywater treated by physical/chemical processes is one of the most frequently questioned and disputed topics across the literature (De Gisi et al., 2016).
Recent risk analyses associated with greywater on-site irrigation concluded that a well-designed treatment system is required for safe greywater reuse (Maimon et al., 2014). The study also called for a more robust exposure estimation for greywater irrigation in home gardening practices (Maimon et al., 2014). However, the risks from neither consuming homegrown food-crop nor toilet flushing using greywater were investigated previously. Considering both practices are feasible and convenient household applications, we carried out a quantitative microbial risk assessment (QMRA) of greywater on-site reuse for toilet flushing and home garden food-crop irrigation to promote the safe reuse of greywater. The risks were quantified and the implications were discussed.

Materials and methods
QMRA was carried out following the classical framework that consists of hazard identification, exposure assessment, dose-response assessment and risk characterization (National Research Council, 1983). The Monte Carlo simulation was used to build a probabilistic-based risk model, so that the range and likelihood of the risk were assessed quantitatively. All calculations were conducted using MATLAB R2017a (The MathWorks Inc., Natick, MA).

Pathogenic E. coli concentration in greywater
Pathogenic E. coli are identified as etiology of various human gastrointestinal illnesses due to the presence of specific colonization factors, virulence factors and pathogenicity associated genes (O'sullivan et al., 2007). Six pathotypes of such strains are now recognized: Verocytotoxigenic E. coli, Enterotoxigenic E. coli, Enteroinvasive E. coli, Enteropathogenic E. coli, Enteroaggregative E. coli (EAEC) and Diffusely Adherent E. coli.
Due to the complexity of methods for quantitative determination of pathogenic E. coli, direct measurements of their concentration in greywater were rare in previous studies. Only O'Toole et al. (2012) reported the detection of virulence gene markers among E. coli isolates but gave no concentration of pathogenic E. coli. In consideration of the data availability, data of total E. coli concentration in domestic greywater were collected instead. A pathogenic ratio was introduced to estimate concentration of pathogenic E. coli based on total E. coli data using, where C PEC is the estimated concentration of pathogenic E. coli in domestic greywater (CFU/100 ml), C EC is the measured concentration of E. coli in domestic greywater (CFU/100 ml), and R path is the pathogenic ratio from E. coli to pathogenic E. coli (unitless). The pathogenic ratio was calculated as the proportion E. coli that are positive for target toxin genes in all E. coli isolates tested according to O'Toole et al.'s (2012) result. In view of the uncertainty of the estimation, the worstcase scenario (R path = 1), in which all E. coli detected were assumed to be pathogenic, was also taken into account through the risk assessment.

Probability distribution fit for E. coli data
Data of E. coli concentration in greywater from different domestic sources are presented in Table 1. As for greywater from bathroom, means and standard deviations of E. coli concentration, in units of CFU/100 ml or log 10 CFU/100 ml, were obtained from six previous studies. E. coli concentration data from each study was assumed to follow a unimodal log 10 -transformed normal distribution because most microbial and environmental measurement data are distributed lognormally (Hirano et al., 1982;Loper et al., 1984).
The log 10 -transformed concentrations (Table 1) were adopted to build a multimodal normal distribution for bathroom greywater, in which sample number of each study was regarded as weight value. The distribution is shown in Fig. 1a, and the fitting parameters are presented in Table 2.
Similarly, the E. coli data of laundry greywater from previous studies were used to build the probability distribution of E. coli concentration (Fig. 1b), except those from O'Toole's study with wash water and rinse water separately reported. A bimodal log 10 -transformed normal distribution was used to integrate these data before they were combined with those from other studies. The volume for each step of laundry was used as weight value (Table 2).
Only four observed values were available for greywater from kitchen (Table 1). Due to the deficiency of data and the existence of various food residues in kitchen greywater, which could lead to considerable uncertainty (De Gisi et al., 2016;Friedler, 2004), a log 10 -transformed uniform probability distribution was applied to estimate E. coli concentration (Fig. 1c). The minimum and maximum values were adopted as two boundaries (Table 2).

Removal rate by treatment process
To account for the likely on-site treatment before reuse, a physical treatment processmicrofiltration was selected as the treatment process for greywater reuse in this study because of its pervasiveness, simplicity and low cost. A 4-log 10 E. coli reduction value was allocated to the microfiltration according to previous experimental results on the removal rate for E. coli (Till et al., 1998;Zheng et al., 2005).

Toilet-flushing scenario
The main ingestion route of pathogenic E. coli through toilet-flushing scenario is inhalation of splashed greywater aerosols. Since pathogenic E. coli mainly cause gastrointestinal infection, only aerosols led to gastrointestinal tracts were taken into account. An assumption was made in this study that all aerosols trapped by human noses are cleared to gastrointestinal tracts to represent a worst-case scenario. This assumption is based on previous reports of human breathing pattern and pathogen ingestion mode through aerosols (Couch et al., 1966;Fry and Black, 1973;Stuart, 1984).
Concentration of aerosols in various diameter sizes at different heights in the air after each toilet flush were measured by O'Toole et al. (2009). Data collected at a sampling height of 420 mm above toilet, after a full flush (9 L) were adopted for the analysis. The sampling height here represents a reasonable scenario that an adult is bending down to flush the toilet or it's a small child who is flushing the toilet.
The inhalation efficiency of aerosols was considered on basis of individuals' breathing pattern during light activities (Moya et al., 2011). A breathing rate of 15 l of air/min, obtained from a breathing cycle period equals 8 s (4 s each for inspiration and expiration) and a 1 -l of tidal volume for each cycle, was adopted. The deposition efficiencies of aerosols in extrathoracic (nasal and laryngeal) region were derived from Heyder et al.'s (1986) study, which was reported as a function of particle size and breathing patterns. Furthermore, a distinction was made between nasal and oral breathing, as they could result in totally different deposition efficiencies of aerosols in extrathoracic region (Stuart, 1984). The exposed duration that refers to time spent in toilet room after one flushing, was set at a typical value of 1 min and a worst-case value of 5 min to represent diverse situations. All values and references of parameters mentioned above are listed in Table 3. The dose of pathogenic E. coli (Dose PEC, toilet ) inhaled and deposited in human gastrointestinal system (in CFU/flush) after each toilet flushing was estimated as: where C PEC, treated is the concentration of pathogenic E. coli in treated greywater (CFU/100 ml), AerosolDose diam i is the mass of water aerosol (according to median diameter size, i) deposited in extrathoracic region (g/min), MFR air is the mean flow rate of air breathed after toilet flushing (l of air/min), T toilet is the time spent in the toilet room after each toilet flushing (min/flush), log MF is the log 10 reduction rate of pathogenic E. coli by microfiltration (unitless), C aero, diam i is the concentration of aerosols in the air splashed after toilet flushing (# of aerosol/-l of air), V aero, diam i is the volume of spherical aerosol (100 ml/aerosol) and DE diam i is the deposition efficiency of aerosols in extrathoracic region (unitless).

Food-crop irrigation scenario
The transfer of pathogens from treated greywater to human body happens when greywater irrigated edible portion of home produce are eaten raw. Such food crops are well-recognized vectors for foodborne diseases (Berger et al., 2010;Olaimat and Holley, 2012), including salad greens, tomatoes, lettuce, cucumber, and pepper. Lettuce was modeled as the representative vegetable in this study due to its popularity as salad green and its propensity to cause human diseases from surface contamination (Lim and Jiang, 2013).
Water retention rate on the surface of lettuce determines the intake rate of contaminants carried in greywater when lettuce is ingested raw. The water adsorption on lettuce using laboratory experiments reported by Shuval et al. (1997) was adopted in the exposure model (Table 3). Water sprays generated during irrigation may represent another pathogen inhalation pathway. However, due to the low exposure volume through water spray inhalation in comparison to food ingestion, the inhalation volume was not included in the assessment. This assumption is further justified by the low probability of human contact during irrigation because most of the spray irrigation in the U.S. occurs at night or a Data of kitchen greywater was not available, the greater value of bathroom and laundry greywater was adopted as an estimation. b Volumes were not available in O'Toole's study, so a study conducted in the same country was referred to.
early morning hours to reduce transevaporation (National Academies of Sciences and Medicine, 2016). It's assumed that home irrigated lettuce is watered every three days, and the environmental decay of E. coli deposited on surface of lettuce occurs between adjacent irrigations. Therefore, a uniform distribution with boundaries of 0 and 3 days was used to estimate withholding time between last irrigation and consumption of lettuce. The inactivation rate of E. coli on lettuce was derived from Sjølander's (2012) study on wastewater irrigated vegetables, which represented a very similar scenario to that used in this study.
The daily intake of lettuce was calculated as a product of human body weight and lettuce intake rate, where the intake rate was expressed as grams of lettuce per kg body weight per day (g of lettuce/kg/day). Empirical distributions of lettuce intake rate and US population' body weight were established from percentile values of survey data reported by previous research (Kahn and Stralka, 2009;Moya and Phillips, 2001) (Table 3).
The dose of pathogenic E. coli (Dose PEC, foodcrop ) ingested through intake of raw lettuce (in CFU/day) was estimated as where log decay is the log 10 environmental decay rate of E. coli on lettuce (log 10 /day), T withhold is the duration of environmental decay (days), R lettuce is the mass of raw lettuce intake per unit body weight per day (g of lettuce/kg/day), M body is the body weight of U.S. population (kg), and V retention is the volume of water retained on per unit weight of lettuce (100 ml/g of lettuce).

Dose-response assessment
The infection or illness risk is commonly expressed as per person per day based on the dose of daily exposure to pathogens. It should be noted that, however, infection includes cases with either symptomatic (showing clinical signs of illness) or asymptomatic (not showing clinical signs of illness) features, while illness only refers to symptomatic cases.
Given a known dose of pathogen, dose-response models, which are generated based on clinical trial data, were used to estimate the risk of a response (e.g. infection or illness). As for pathogenic E. coli, the most widely accepted dose-infection model is characterized as a beta-Poisson model (DuPont et al., 1971), where P inf is the estimated infection risk, Dose represents the dose of pathogenic E. coli ingested (CFU), α and N 50 are best-fit parameters of the model (Table 4).
The dose-illness model is characterized by an exponential function (DuPont et al., 1971),  where P ill is the estimated illness risk, and k is the best-fit parameter of the model which represents the pathogenicity of pathogenic E. coli.
The parameter values used in dose-response model are listed in Table 4.

Risk characterization
The acceptable annual infection risk level proposed by the EPA (2005) and the acceptable disability-adjusted life years (DALYs) proposed by WHO (2008) are two most authoritative and widely-used health risk benchmarks. The U.S. EPA benchmark is ≤10 −4 infection cases pppy, while the WHO benchmark is ≤10 −6 DALYs pppy.
The annual infection and illness risk were calculated based on the theorem of independence using (Haas et al., 1999)  where P inf, annual scenario and P ill, annual scenario are the estimated annual infection risk and annual illness risk, Freq scenario is the number of times a certain scenario occurs during a day, and n represents the total number of occurrence of the scenario in a year.
For toilet flushing scenario, a frequency of 8 times a day was applied to represent the worst-case scenario of a healthy human. The frequency of consumption of greywater irrigated lettuce was set to once a day, because the Dose PEC, foodcrop described in Section 2.2.2 represents the dose of pathogenic E. coli ingested through intake of raw lettuce per day.
For annual illness risk, the DALYs was calculated as (Mara et al., 2007) where DALYs illness case PEC is the disability-adjusted life years per illness case caused by pathogenic E. coli (Havelaar et al., 2015) (Table 4).

Monte-Carlo simulation
Monte-Carlo algorithms were written to estimate the probability distribution of microbial risks in both scenarios. Each input parameters (e.g. E. coli concentrations, concentration of aerosols in air after each toilet flushing and lettuce intake rate) were randomly selected from their probability distributions. The pseudo-algorithm information flow is shown in Fig. 2. Output parameters (e.g. doses of pathogenic E. coli ingested, annual infection risks and disease burdens) were computed over 10,000 iterations so that the distributions can reach a steady state. Small variation (i.e. b1%) in terms of average between replicates of distribution was used for checking reproducibility of model outputs (Lim and Jiang, 2013).

Sensitivity analysis
A local sensitivity analysis was used to assess the variability propagation of each input parameters throughout the risk models (Gottschalk et al., 2010a;Gottschalk et al., 2010b;Manheim and Jiang, 2017). The true means of distributions (or the values of pointestimates) were adopted as baseline point values for each input parameter and output variable. Then a differential value for each output variable X mean was calculated by decreasing the baseline input parameter P mean value by 10% (Gottschalk et al., 2010b). The sensitivities of annual infection risk and disease burden related to each input parameter (e.g. E. coli concentration, concentration of aerosols, water retention rate, etc.) were calculated as where S is the sensitivity value (unitless), X mean is the mean of the output variable distribution using the original values, ΔX mean is the difference in means between the original output distribution and the changed output distribution, P mean is the mean of the original input distribution (or the value of the original point-estimate), and ΔP mean is the difference in means between the original input distribution and the changed input distribution (or the difference between values of original and changed point-estimates). In addition, to identify the most influential contributors to the predicted health risks, all sensitivity values for input parameters were summed to calculate their relative contribution to the total sensitivity.

Toilet-flushing scenario
As shown in Fig. 3, infection risks of pathogenic E. coli through toilet flushing using treated domestic greywater are almost negligible. For all scenarios discussed, the infection risks (median range: 8.8 × 10 −15 -8.3 × 10 −11 , 95th percentile range: 2.2 × 10 −11 -4.8 × 10 −8 ) are significantly less than the U.S. EPA annual infection benchmark (≤10 −4 pppy) by orders of magnitude. Even in the worst-case scenario, where all E. coli are assumed to be pathogenic, the infection risks are still far below the benchmark.
The infection risks from bathroom greywater and laundry greywater are very close to one another, while the infection risk from kitchen greywater exhibits much higher risk than the former two by approximately two orders of magnitude. The breathing style, as revealed in Fig. 3, plays a similar role, where infection risks for nasal breathers are about 2 log 10 higher than those for oral breathers. Situations with different durations spent in toilet room after a flush are not specially discussed, because the difference for risk between one-minute and five-minute exposure is within one order of magnitude (Appendix B, Table B.1).

Food-crop irrigation scenario
Infection risks of pathogenic E. coli from consuming food-crops irrigated by treated greywater (median range: 2.6 × 10 −8 -4.9 × 10 −6 , 95th percentile range: 1.0 × 10 −4 -4.5 × 10 −3 ) are much closer to (or even exceed) the U.S. EPA benchmark (Fig. 3) in comparison with toilet flushing scenarios. Similar to the former application, the infection risk from kitchen greywater is about 2 log 10 higher than those from bathroom or laundry; the median risk of kitchen greywater (value: 1.7 × 10 −4 ) slightly goes beyond 10 −4 pppy in the worst case and poses a considerable threat for human consumers (Appendix B, Table B.2).

Sensitivity analysis
The relative contribution of each input parameters to the variability of infection risks and disease burdens are summarized in Fig. 5. Among the model inputs included in toilet flushing scenario, the output annual infection risks are most sensitive to the input E. coli concentration (fraction range: 59.22%-61.51%). Time spent in toilet room (T toilet ) and the daily frequency (Freq toiletflushing ) have the same contributions in each scenario and are far less than that of E. coli concentration. Concentrations of aerosols (C areo, diam 1 , C areo, diam 2 ) are most irrelevant inputs with a combined fraction b10%.
In food-crop irrigation scenario, the input E. coli concentration also contributes the highest sensitivity fractions for the output annual infection risks. Especially in the scenario of kitchen greywater reuse, the fraction is nearly 80%. Other input parameters all represent minor contributors to the variability of the infection, with fractions b10%.

Implications
The results of the study showed that greywater food-crop irrigation exhibits a higher health risk than toilet flushing under same conditions. Different greywater sources also lead to disparities between estimated risks. Treated greywater from all three household sources can be used for toilet flushing without significant health risks, while for food-crop irrigation scenario, only greywater from bathroom and laundry can be reused within acceptable risks.
The infection site of the target pathogen may partly explain the low risks associated with toilet flushing scenario. Pathogenic E. coli mainly causes gastrointestinal infection, which means the pathogens have to go through the respiratory tract before they ultimately enter their infection site, the digestive tract. Under such condition, only greywater aerosols intercepted in extrathoracic region are possible to be ingested, which result in low exposure doses, and thus low relevant risks. The reason that nasal breathers exhibit a higher risk in this scenario is because of the higher interception efficiency on aerosols. This might be a consequence of the worst-case assumption that all aerosols trapped in extrathoracic region are led to gastrointestinal tracts, since nasal breathing is often considered to be healthier. Future research that provides better understandings of the pathogen ingestion through aerosol inhalation will improve the confidence in this estimate.
Annual infection risk and disease burden are adopted as indicators throughout the analysis. When calculating disease burdens, the unique characteristics of different pathogens on morbidity and mortality are taken into account, which means a pathogen with higher virulence will get a greater DALYs per illness case. In contrast, analyses with annual infection risk regard all pathogens the same important (Gibney et al., 2013). Difference between two indicators could cause sharp disagreement on discussion of risk acceptability.
In this study, only one target pathogen (pathogenic E. coli) is used to estimate the health risks, which means all DALYs are calculated by the same dose-illness model. This approach results in the same trends between annual infection risks and disease burdens under the same condition. However, DALYs indicate lower threats than annual infection risks in each scenario according to corresponding benchmarks. This is likely due to the relatively low virulence of pathogenic E. coli. It should be noted that, both the WHO's benchmark of ≤10 −6 DALYs pppy and the U.S. EPA benchmark of ≤10 −4 infection pppy are established for assessing safety of drinking water. And both are considered overly conservative and impractical for the health risk assessment on non-potable water uses (Mara, 2011;Mara and Sleigh, 2010).

Concentration of the target pathogen
Pathogenic E. coli is one of the many pathogens in the greywater that may pose considerable risks. Lack of quantitative pathogen data in greywater has been the major hurdle for a comprehensive risk analysis. Probability distribution of E. coli concentration in greywater is established based on measured data from various regions of the world ( Table 1). The data mining results indicate that concentrations of E. coli in bathroom and laundry greywater vary little despite the geographic differences, which certain their reliability in risk estimation in the U.S. In contrast, qualities of kitchen greywater in different societies can vary significantly due to different living habits. For example, Americans are used to directly discarding food residue into kitchen sinks after meals, while Chinese people only use the kitchen sink for washing vegetables and rinse dishes. These household practices can cause considerable difference between kitchen greywater qualities. Since E. coli concentration has the greatest influence on risk outcomes based on the sensitivity analysis, additional data on pathogen concentrations in kitchen greywater, especially the local ones, are needed for a more accurate and more reliable health risk assessment.
The pathogenic ratio of E. coli used in the model is only calculated by a positive-or-negative detection instead of exact concentrations of pathogenic and total E. coli (O'Toole et al., 2012). The credibility of ratio value may contribute to a remarkable uncertainty, the range of which is considered in this study by discussing the worst-case scenario with all E. coli being pathogenic and is reducible when improved knowledge available.

Toilet flushing scenario
During toilet flushing, the aerosols generated are proportional to the amount of energy and water volume used in a single flush (Johnson et al., 2013). O'Toole et al. failed to detect any aerosols after flushes with flush volumes of below 4.5 L (O'Toole et al., 2009). Present water-saving toilets are commonly equipped with a full flush volume of 6 L and a half flush of 3 L, which will produce lower aerosol volume and health risks in comparison with the 9 L full flush toilet modeled in this study. Johnson et al. also reported that the concentration of aerosols drop with the increase of sampling height (Johnson et al., 2013). The aerosol concentration at a location 420 mm above the toilet seat described in Section 2.2.1 is much lower than the common height of a standing adult's nose (or mouth). Considering the short duration of a person to bend down to flush the toilet, aerosols exposed to human noses are likely at lower concentrations than the conservative assumption used in this study. Furthermore, gravitational shrinkage or sedimentation of aerosols mostly happens within the first 30s after a single flush (Johnson et al., 2013). The dynamic reduction of aerosol concentrations in the air is not included in the model since data are not available. The initial aerosol concentrations are used for the entire exposure as a worst-case estimation. Other factors such human breathing patterns using mix nasal and oral breathing can also influence the risk outcome as indicated in a previous study (Lim et al., 2015).
After all, the aerosol concentrations and the duration of breathing polluted air are not sensitive input parameters for the model. Based on the sensitivity analysis, the model predictions can be afforded relatively high levels of confidence, despite the uncertainties discussed above.

Food-crop irrigation scenario
The health risk result of food-crop irrigation scenario is only fully applicable to the U.S. population because data of body weight and lettuce intake rate used in the model are generated from American domestic surveys. However, since the model outcomes are insensitive to neither of these two parameters, the results are still referable in other regions. Other uncertainties may include pathogen remove from washing lettuce before consumption. In the absence of credible data, a worst-case Fig. 6. Estimated hourly patterns of greywater production and cumulative volume. scenario of no washing before eating is discussed to represent the highest potential risk. Despite that, a thorough washing before eating is highly recommended to dilute possible pathogens retained on produce surface, thus effectively reduce the related health risks (according to the high sensitivity of E coli. concentration).

Dose-response models
The most important source of uncertainties, perhaps, is the doseresponse models used for estimating the infection or illness probability through a single exposure. Of six known pathotypes of E. coli, only EAEC and Enteropathogenic E. coli were detected in bathroom greywater and in laundry greywater, respectively (O'Toole et al., 2012). However, due to the lack of relevant studies, pathotypes of E. coli in greywater from different sources can't be simply defined. Although the dose-response models of different pathotypes are not exactly the same due to different infection or illness mechanisms, dose-response models have not been established for all (e.g. EAEC) (DuPont et al., 1971;Feeguson and June, 1952;Graham et al., 1983;Haas et al., 1999;June et al., 1953;Levine et al., 1977). All pathotypes of E. coli are considered the same, as 'pathogenic E. coli', in the study; and two most widely accepted doseresponse models (dose-infection model and dose-illness model) on pathogenic E. coli are used for estimation (Enger et al., 2013;Hayashi, 2016;Weir et al., 2017;Wilkes et al., 2013). A more accurate risk estimation calls for more field study of pathogenic E. coli in greywater and clinical infection data of various pathogenic E. coli pathotypes.

Contribution and limitation
The interpretation of QMRA results is usually made through comparison with existing water quality standards. For instance, California has adopted a strict microbial standard for toilet flushing using reclaimed water, which requires a 7-day median concentration of ≤2.2 total coliforms/100 ml of water (Lim et al., 2015). Similarly, the standard in China specifies the number of total coliform below 3 per liter of water, which is magnitudes lower than those of German (100 total coliforms/ 100 ml of reclaimed water) and Japan (1000 total coliforms/100 ml of reclaimed water) (De Gisi et al., 2016). These rules are formulated on basis of different purposes and considerations, but they all intend to make simple yes-or-no judgments by only one (or two) commonly adopted microbial indicator(s) (De Gisi et al., 2016). These over simplified policy decisions are primary due to the consideration of ease for implementation and supervision.
In comparison with applying numerical E. coli standards to treated greywater quality, QMRA incorporates type and source of pathogens in specific greywater stream and exposure scenarios to provide the probability distributions of the infection/illness risks. The results of QMRA, which is often regarded as a more pertinent approach for risk characterization, can serve as scientific basis for revision of those standards.
In this study, microfiltration is selected as a typical process to simplify the model. The low health risk of greywater treated by microfiltration can also be generalized for other commonly used membrane processes with smaller pore sizes, such as ultrafiltration, nanofiltration and reverse osmosis. Furthermore, health risks of other treatment processes, such as disinfection used in Aquasave Project (ENEA, 2002), can be estimated using the similar approach, by simply inputting their removal rates (either point estimate or probability distribution) instead of that of microfiltration.
It should be noted that, however, greywater treatment requirement for on-site reuse is beyond health risk concern. For instance, the fouling and odor caused by organics can seriously hinder the practicality of using greywater for toilet flushing (Kuru and Luettgen, 2012). Another threat comes from inorganic salts or surfactants in greywater that might pose great environmental risk to ecosystems through the reuse for irrigating gardens (Lubbe et al., 2016;Pandey et al., 2014). Other basic organoleptic indicators, including color and turbidity, should also be taken into consideration for improving public acceptance of greywater reuse.
Furthermore, organics and other nutrients in greywater, if not well eliminated, may give rise to regrowth of microorganism when the water is stored in tanks (March et al., 2004). According to the latest data on volumes (DeOreo et al., 2016) and patterns (Mayer et al., 1999) of indoor water uses in a single U.S. household, the hourly greywater production (excludes water from kitchen sink faucets) commonly exceeds the water consumption of toilet flushing (Fig. 6). Assuming all stored greywater in tank are used for outdoor irrigation or drained into the sewer at 6:00 and 18:00 every day (Mayer et al., 1999), there are still greywater accumulation in storage tanks between evacuation.
However, the dynamic process that greywater getting in and out the storage tank makes the estimation of pathogen regrowth far more complicated than a simple time-based growth function. It should also be noted that, only regrowth of indicator microbes (e.g. E. coli) was reported in greywater tanks (March et al., 2004), which doesn't necessarily refer to regrowth of real pathogens (e.g. pathogenic E. coli). In fact, treated domestic greywater is not often regarded as good environment for pathogen growth. Additional studies are needed to elucidate pathogens' behavior during greywater storage.

Conclusions
The risk assessment outcomes of using treated household greywater indicate: • Given the same greywater source, food-crop irrigation exhibits a higher health risk than that of toilet flushing under same conditions. • Given the same exposure scenario, kitchen greywater poses the highest risk, followed by laundry greywater and bathroom greywater. • Greywater from bathroom and laundry are safe for both toilet flushing and food-crop irrigation after treated by microfiltration. • Treated greywater from kitchen is not clean enough for food-crop irrigation while it's innocuous for toilet flushing. • Many factors contribute to the uncertainties of the risk outcomes.
Among them dose-response model and pathogen concentration are the most critical at accuracy of the estimates. • Overall, greywater on-site reuse should be promoted with proper awareness of the risk.