**1. Introduction**

There is growing evidence that air pollution and specifically fine particulate matter (PM2.5) contribute significantly to health burden and further, there is a close relationship between long-term air pollution exposure and adverse health e ffects in urban populations [1,2]. The assessment of Global Burden of Disease (GDB) indicated that PM2.5 contributed 4.24 million deaths globally in 2015 [3]. Assessments of human health e ffects attributed to an air pollutant are dependent on the magnitude of human exposure to that pollutant. Thus, the accuracy of a health burden assessment is determined by the uncertainty of predicted population exposure. Quantifying the population exposure to air pollution is subject to several challenges:

The spatiotemporal variability of ambient concentration is strongly influenced by emissions dynamics, predominantly from road transport, (such as peaks in tra ffic-related pollution during rush hours), meteorological conditions, which determine the transport and dilution of air pollutants and local conditions such as the urban form (e.g., the presence of high buildings can reduce the dispersion of the pollutants), which are the most important factors leading to significant variation of air pollutants in urban areas.

The proportion of outdoor air infiltrated to indoor microenvironments (MEs) is influenced by di fferent housing designs and patterns of behaviour inside the building.

The spatiotemporal variability of people's activity (population time–activity patterns) in various MEs [4].

Around 75% of European populations live in cities, with a highly variable range of activities carried out at di fferent times and in di fferent places [5]. The quality of data, or the absence of key components within an epidemiological exposure assessment, is likely to a ffect the magnitude and significance of the prediction misclassification in a health burden assessment (Figure 1).

**Figure 1.** Schematic diagram of an exposure assessment structure for health burden misclassification.

Traditionally, epidemiological studies relied on centralized ambient concentration measurements of limited monitoring sites [6–10]. This is likely to lead to an exposure error, since several monitoring studies have suggested that air pollution data from a single site can represent only a small surrounding area especially in urban environments, due to pollutants' spatial heterogeneity [11,12]. Ambient air pollutant concentration can be estimated in several ways such as through field observations, statistical modelling such as land-use regression (LUR) and air quality dispersion models (AQM) that can use various spatial resolutions [13]. Willers et al. [14] indicated that using air quality data measured at a single site and assuming that exposure across cities was the same, could cause considerable misclassification of exposure. In their study, they examined the di fference in mortality risk between neighborhoods in the city of Rotterdam and found that the mortality risks between neighborhoods had a di fference of up to 7%. By utilizing land use regression techniques and air quality models, several studies have managed to demonstrate that an increased spatial resolution of the exposure concentration could lead to significantly di fferent exposure or health burden estimates [15–18]. Similarly, Punger and West [19] assessed the e ffect of spatial resolution to population-weighted PM2.5 concentrations in the

U.S. by utilizing the Community Multiscale Air Quality (CMAQ) model. They found that population exposures, maximum concentrations and standard deviations all reduced at coarser resolutions. At 408 km resolution, exposure and maximum concentration were 27% and 71% lower, respectively than those at 12km resolution. Attributable mortality also reduced as the resolution became coarser. Several studies have shown that coarse resolutions might result in lower mortality attributed to PM2.5 [20]. Fenech et al. [21] concluded that total mortality estimates were sensitive to model resolution up to ±5% across Europe, whereas Korhonen et al. [22] found that, considering only local sources of primary PM2.5, the mortality reduced by 70% in the whole country (Finland) and 74% in urban areas when the resolution changed from 250 m to 50 km.

Apart from the exposure misclassification due to the di fferent levels of spatiotemporal resolution of outdoor concentration, there are other significant contributors, in particular the infiltration of outdoor pollutants to indoor MEs and di fferent time-activity patterns in MEs. As particles infiltrate and persist indoors, where people living in urban areas spent over 80% of their daily time [23], most of the exposure to PM2.5 actually occurred in the indoor microenvironments [23–25]. The fraction of ambient PM2.5 that infiltrates indoor microenvironments can vary due to particle size, building characteristics, meteorological conditions and human activities [26]. Consequently, relying on outdoor measurements alone can therefore lead to exposure misclassification. Moreover, variations in the time spent in various MEs (e.g., outdoors, indoors, vehicles, subway) also influence population exposure to outdoor-generated PM2.5 due to the spatial variability of both outdoor concentrations and the indoor transport of ambient PM2.5. Baxter et al. [27] compared four di fferent approaches to PM2.5 exposure prediction, where each model was of a di fferent complexity. In their study they focused on the heterogeneity in exposures but did not investigate the influence on health e ffect predictions. They suggested that geographic heterogeneity in both housing stock (and thus a relatively consistent Air Change Rate) and human activity patterns contribute to significant heterogeneity in ambient PM2.5 exposure both within and between cities that is not demonstrated by stationary monitors. Ma et al. [28] compared three di fferent types of PM2.5 exposure estimates to illustrate the di fferences in exposure levels between estimates obtained from di fferent approaches. They found that the daily average PM2.5 exposures for residents with di fferent activity patterns may vary significantly even when they were living in the same neighborhood. Several studies have also investigated the correlation between outdoor PM2.5 and mortality, although their results are skewed by the fact that people spend the majority of the time indoors. Ji and Zhao [29] used existing epidemiological data on ambient PM2.5-related mortality to estimate mortality associated with indoor exposure to outdoor-generated PM. This was the first attempt to quantify that relationship and their results indicated that outdoor PM had substantial e ffects on health caused by exposure within indoor MEs. Recently, Fenech and Aquilina [30] used the annual mean PM2.5 concentrations derived from local fixed monitoring stations to estimate the PM2.5-related mortality in the Maltese Islands. They found that the attributable fraction of all-cause mortality associated with long-term PM2.5 exposure ranged from 5.9% to 11.8%, indicating that PM2.5 concentration is a major component of attributable deaths. Azimi and Stephens [31] used a modified version of the common exposure-response function and developed a framework for estimating the total U.S. mortality burden attributed to exposure to PM2.5 of both indoor and outdoor origins. They found that residential exposure to outdoor-generated PM2.5 accounted for 36% to 48% of total exposure, indicating that e fforts to mitigate mortality associated with exposure to PM2.5 should consider indoor pollution control as well.

That of particular importance is how di fferent exposure approaches impact long-term health burden/mortality predictions and the magnitude of the resultant impact. We made multiple comparisons between refined ambient PM2.5 exposure surrogates (that account for important factors such as the infiltration and time-activity) and the fixed-site monitor PM2.5 concentrations to indicate the importance of including more dynamic data to epidemiological studies and to demonstrate how more complex modelling approaches modify mortality predictions. By using BenMap-CE we were able to provide the spatial distribution of health outcomes influenced by the exposure misclassification. While a number of studies have already investigated exposure misclassification when using different approaches and others have estimated health effects based on specific exposure metrics, the aim of this work is to move one step further and answer the question: how much is the misclassification that occurs when using different exposure approaches to predict health burden?

#### **2. Materials and Methods**

This work aims to quantify the long-term health burden misclassification that occurs when different PM2.5 exposure metrics are utilized. An ecologic design was used to generate associations between air pollution exposure and health outcomes. We investigated the Greater London Area (GLA), building on recent exposure studies that have explicitly estimated London population exposure using hybrid dynamic models [32]. Here, we have described five different exposure Tier-models of incrementally increased complexity are considered by gradually including data of important MEs, such as infiltration rates of the different dwelling types and the London Underground, where London's population spend most or part of their daily time. The London Travel Demand Survey (LTDS) space-activity data were categorized into three major ME groups. The analysis estimated the magnitude of the change (i.e., avoided or incurred) in mortalities when moving from the central-site monitored concentrations as a surrogate for population exposure (Tier-model 1) to more refined exposure Tier-models. The original ambient PM2.5 concentrations were based on average hourly data measured by 23 monitoring stations located in the GLA [33] and the examined MEs were: i) indoors (i.e., home-indoor), ii) aboveground transportation iii) the London Underground and iv) outdoors. The following sections describe the structure of the methodology and the development of each component.

#### *2.1. Developing Tier Models to Estimate Human Exposure*

To capture different exposure assessment methods that have been used in epidemiology, we developed five different Tier models of increased complexity, moving from static to more dynamic approaches (Table 1). This method was separated into two parts: i) The microenvironments and time-activity patterns were classified and calculated based on the derived information; ii) the time-activity information was matched with corresponding microenvironmental concentrations to estimate the dynamic time-weighted exposure. The exposure time was considered costly and the metrics estimated the annual hourly-average PM2.5 exposures, which were then used as an input for BenMap-CE [34].


**Table 1.** Tier models for assessing the time-weighted exposure.

The Tier-model stages and the respective approaches are briefly described below. Tier model 1: Outdoor

$$E = \mathbb{C}\_{\text{out}} \tag{1}$$

where E is mean exposure and Cout is mean outdoor concentration of PM2.5.

Hourly readings were extracted from the London Air Quality Network (LAQN) [33]. LAQN consists of automatic monitoring equipment in fixed cabins, which measures air pollution at breathing height. It provides electronically available data on concentrations of major urban pollutants and has been used in several studies [35,36]. The ratified concentration data from 23 available monitoring stations in GLA were downloaded and added to BenMap-CE. Only the monitors that could provide at least 70% of the data for the whole year were selected. The ambient concentration was considered as representative of the total population exposure.

Tier model 2: Indoor

$$\mathbf{E} = \mathbf{C}\_{\text{in}\prime} \tag{2}$$

where Cin is the mean indoor (i.e., home-indoor) concentration.

This Tier model utilized the information of the spatially distributed concentration and the total average Indoor/Outdoor (I/O) ratio in GLA to estimate the exposure inside the residence [37].

Tier model 3: Indoor (dwellings)

$$\mathbf{E} = \sum \mathbf{C}\_{\text{out}} \mathbf{\*} \, \mathbf{F}\_{\text{i}} \, \mathbf{x}\_{\text{i} \prime} \tag{3}$$

where Fi is the infiltration rate of each dwelling type (i) and xi is the frequency (%) of this type in London.

In this study, all the indoor environments were combined into one single ME (i.e., home-indoor) without considering other indoor environments, such as office or commercial buildings, due to the lack of infiltration data. Subsequently, the I/O ratios that we used also represented offices and other indoor places, assuming that the I/O ratios for other indoor MEs had the same values as domestic home buildings [32]. The I/O ratios of London's housing stock were obtained from Taylor et al. [37]. In their study they estimated the Indoor/Outdoor ratio of 15 building archetypes. We grouped these archetypes into five main dwelling types in response to available housing stock data in Middle-Super-Output-Area resolution obtained from the Mayor of London, Datastore [38]: i) flat, ii) bungalow, iii) terraced, iv) semi-detached and v) detached (Table 2). The frequency of each type could be calculated from the number of properties in the GLA, which represented 98.7% of the housing (The average I/O ratio was assigned to the unknown 1.13%). Figure 2 shows the annual average I/O ratios of PM2.5 concentration in the GLA. The average ratios, including all dwelling types and their frequency, ranged from less than 0.54 to 0.59. The highest ratios were observed in Outer London, whereas the lowest ratios were observed in Inner and South West London, probably due to the newer building stock and the large number of flats in large buildings (London Datastore), where the available surface for infiltration was considerably smaller.

**Table 2.** London's dwelling group type descriptions, frequency in stock and average Indoor/Outdoor (I/O) ratios.


Tier model 4: Outdoor + Indoor + Transportation (aboveground and underground)

$$\mathbf{E} = (\mathbf{C}\_{\text{out}}\,\mathrm{\*}\,\mathrm{t}\_{\text{out}}) + (\sum \mathbf{C}\_{\text{out}}\,\mathrm{\*}\,\mathrm{F}\_{\text{i}}\,\mathrm{\*}\,\mathrm{x}\_{\text{i}})\,\mathrm{\*}\,\mathrm{t}\_{\text{ind}} + (\sum \mathbf{C}\_{\text{out}}\,\mathrm{\*}\,\mathrm{F}\_{\text{j}})\,\mathrm{\*}\,\mathrm{t}\_{\text{abg}} + (\mathbf{C}\_{\text{undg}}\,\mathrm{\*}\,\mathrm{t}\_{\text{undg}}),\tag{4}$$

where (j) is each aboveground transport-ME (tMEs) and tout, tind, tabg and tundg is the fractional time spent (%) annually outdoors, indoors, aboveground tME and London Underground (LU) tME, respectively.

This Tier-model includes transportation as an additional microenvironment, where an urban population spends time during the day. This ME was categorized into aboveground and underground transportation. Aboveground transportation refers to car, bus and train, whereas underground to

London subway. By separating transportation into 2 groups we were able to evaluate the influence of a highly polluted ME, like the London Underground (described in the next section), on the total population exposure concentration.

**Figure 2.** Map of annual average Indoor/Outdoor (I/O) ratios used in our study.

The space–time–activity data for our study were based on the London Travel Transport Agency (LTDS) of Transport for London (TfL) [39] for the period between 2005 and 2010 (Table 3). The data were generated from the interviews of approximately 8000 households per year, providing very useful information about their daily time–activity patterns, including travel modes and trip times. The data were scaled to represent the population of London, excluding children under five years old [32].


**Table 3.** Summary table of the time–activity data.

According to Smith et al. [32], the average daily percentage of time spent indoors was 95.7 %, whereas people spent 2.5%, 0.4% and 1.4% in aboveground transportation, London Underground and outside (walking or cycling), respectively. This proportion of time spent indoors also includes approximately 20% of surveyed people, who did not leave their house. In this study, these percentages were used as annual averages for the whole population over five years old, including the different times spent during weekdays and weekends.

For the in-vehicle exposure of the aboveground sub-microenvironment, we calculated the PM2.5 concentration by solving the mass balance equation [30]:

$$\text{dC}\_{\text{in}} \text{ / dt} = \lambda\_{\text{win}} \, \text{\*} \, (\text{C}\_{\text{out}} - \text{C}\_{\text{in}}) - \eta \lambda\_{\text{HVAC}} \, \text{\*} \, \text{C}\_{\text{in}} - \text{V}\_{\text{g}} \, \text{\*} \, (\text{A}^{\prime} / \text{V}) \, \text{\*} \, \text{C}\_{\text{in}} + \text{Q} / \text{V}, \tag{5}$$

where Cout is the outdoor concentration around the vehicle, Cin the concentration inside the vehicle, λwin and λHVAC are the hourly air exchange rates from the windows and mechanical ventilation system, respectively, n is the filter removal efficiency taking values between 0–1, Vg is the deposition velocity in (m/h), A' is the internal surface area, V is the volume of the vehicle and Q is the in-vehicle particle emission rate in μg/h. To solve this equation, the same values with Smith et al. [32] were used except for the concentrations and the commuter's surface was derived from Song et al. [40], in order to calculate A'.

Tier model 5: Outdoor + Indoor + Transportation (aboveground and underground→ deep lines + subsurface lines).

The time-weighted exposure equation associated with this Tier model stage is:

$$\begin{array}{l} \mathrm{E} = \left( \mathrm{C}\_{\mathrm{out}} \, ^{\ast} \mathrm{t}\_{\mathrm{out}} \right) + \left[ \left( \sum \mathrm{C}\_{\mathrm{out}} \, ^{\ast} \mathrm{F}\_{\mathrm{i}} \, ^{\ast} \mathrm{x}\_{\mathrm{i}} \right) \, ^{\ast} \mathrm{t}\_{\mathrm{ind}} \right] + \left( \sum \mathrm{C}\_{\mathrm{out}} \, ^{\ast} \mathrm{F}\_{\mathrm{j}} \right) \, ^{\ast} \mathrm{t}\_{\mathrm{b}} \mathrm{g} + \left( \mathrm{C}\_{\mathrm{undg}\, \mathrm{h} \, \mathrm{v} \mathrm{ac}} \right) \, ^{\ast} \mathrm{t}\_{\mathrm{undg}\, \mathrm{h} \, \mathrm{v} \mathrm{ac}} \\ \qquad + \left( \mathrm{C}\_{\mathrm{deep}\, \mathrm{v} \, \mathrm{undg}} \, ^{\ast} \mathrm{t}\_{\mathrm{deep}\, \mathrm{v} \mathrm{undg}} \right) . \end{array} \tag{6}$$

In the 5th and most complex Tier model, the same procedure as in Tier 4 was followed but the London underground microenvironment was further divided into subsurface and deep lines to reflect the significant difference in concentration on two types of lines. The use of mechanical ventilation in the subsurface lines results in much lower PM2.5 concentrations than the deep lines due to air filtration (explicitly described in the next section). Hence, by dividing the underground into two subgroups we were able to improve the exposure estimates and to examine the contribution of a very highly polluted microenvironment to the total exposure. The proportion of time spent in each of those two subcategories was assumed according to the number of annual journeys completed in each line during 2017, where 77% were made by the deep-line underground and 33% by the subsurface.

#### PM2.5 Concentration in the London Underground

As the London Underground microenvironment was unable to be accurately represented by the outdoor measurements, due to its high concentration of PM2.5 and its limited connection to the outside world, a series of air pollution measurements were conducted inside the London Underground. The PM2.5 measurements took place on five major London Underground platforms and trains (Bakerloo line, Circle line, Central line, District line and Victoria line) by using the portable DustTrak II Aerosol Monitor 8534, a light scatter laser photometer, which could provide a large number of real-time readings. The current selection of the lines was decided in order for both the deep without mechanical ventilation lines and the subsurface with HVAC lines to be represented by our measurements.

Our original intention was that the measurements would reflect the cold and the warm period of 2017. Hence, the experiment was conducted during the morning and the afternoon for one week in February and one week in July. The average concentration in the London Underground for the whole year was very high, approximately 218 μg/m3, albeit when we grouped the lines into deep without HVAC lines (Central, Bakerloo and Victoria) and subsurface lines with HVAC (Circle, District) we noticed a remarkable difference between the two concentrations (70.2 μg/m<sup>3</sup> for the subsurface lines and 365.6 μg/m<sup>3</sup> for the deep lines). The PM2.5 concentration levels in the unmeasured lines were assumed to be similar to these measured. The classification of the unmeasured lines was made according to their depth and ventilation system.

In the London Underground, Seaton et al. [41] reported higher platform concentrations of 480 μg/m3. Recently, Smith et al. [42] assessed day to day variation in LU concentrations and compared them with those above ground. During their campaign, 22 repeat journeys were made on weekday mornings over a period of five months. They found that the subsurface ventilated District line had the lowest PM2.5 concentration levels (i.e., mean 32 μg/m3) and the deep unventilated Victoria line the highest (i.e., mean 381 μg/m3), while the mean concentration in the LU, according to their measurements, was 302 μg/m3. Although their monitoring method and equipment were different from those used in this study and the sampling period was longer, their findings do not differ significantly from ours. Even though the station measurements in the UK are limited, most of the studies made so far have measured approximately two times higher concentrations in the London Underground than in other undergrounds worldwide [43,44], probably due to its age and the limited ventilation systems.

#### *2.2. Simulating PM2.5 Exposure Concentration and Estimating Health Impact Using BenMap-CE*

The environmental Benefits Mapping and Analysis Program—Community Edition (BenMap-CE) is a powerful Geographical Information system (GIS)-based program that estimates the health e ffects associated with the change in air quality [34,45]. These data consisted of a middle layer super output areas (MSOA) map of GLA, the derived monitoring data and London's population data, in order to estimate the health impact. BenMap-CE provides three interpolation methods: the closest monitor, the fixed radius, and Voronoi Neighbour Averaging (VNA). Among the incorporated methods, VNA was the most suitable for our case, covering the unmonitored areas and giving the best spatial distribution of the concentration.

After uploading the essential data and determining the appropriate Health Impact Function (HIF) for our analysis, we were able to quantify the health impact misclassification (i.e., change in all-cause mortality, either incurred or avoided) resulting from the exposure metric di fferences. In this study, the following long-term health impact function was used to estimate the change in all-cause mortality [46]:

$$
\Delta \mathbf{Y} = \mathbf{Y}\_0 \mathbf{\*} \left( 1 - e^{-\beta \Lambda PM} \right) \mathbf{\*} \text{Pop}\_{\prime} \tag{7}
$$

where ΔY is the change in health e ffect, Y0 is the baseline mortality rate (the mortality rate at minimum risk concentration), β is the unitless beta coe fficient, Δ*PM* is the change in the exposure rates between Tier 1 and the other Tier models (Tier 1 is the base case) and Pop is the exposed population.

One limitation of the aforementioned e ffort to estimate the health impact of indoor air pollution is the use of the mortality e ffect estimate (i.e., beta coe fficient) that is usually taken directly from the epidemiology literature on the studies conducted for outdoor air pollution. Therefore, to account for that fact, some studies on the health e ffects of outdoor-generated PM2.5 introduced a method for modifying the mortality e ffect estimate (i.e., beta coe fficient) based on the average infiltration factor combined with the mean fraction of time spent in indoor MEs [13,47,48]. However, the application of the adjusted coe fficient is solely for the component of indoor PM2.5 of outdoor origin and not of indoor PM2.5 in total. The way indoor particle sources are treated has a larger impact than the adjustment of the coe fficient for the outdoor-generated fine particles and remains an evidence gap of considerable public health importance. In another study, Logue et al. [49] used a central estimate of the beta coe fficient for premature mortality related to both indoor- and outdoor-generated PM2.5, which was directly derived from the epidemiology literature. In our case, due to the mobile monitoring conducted in the LU and the distinct function of BenMap-CE, a central mortality e ffect coe fficient from Pope et al. [50] was used as an input. The mortality e ffect coe fficient was utilized to generate BenMap's health impact functions in the direction of estimating the change in estimates of mortality (either avoided or incurred) when using di fferent exposure metrics. Furthermore, we estimated the percentage decrease in the predicted avoided cases when moving from the less complex (static) metrics to more dynamic metrics.
