The present study developed a framework through which personalised estimates of exposure to ambient air pollution concentrations, accounting for indirect information on time-activity and population mobility derived from a representative sample, could be applied to residence-based estimates traditionally assigned to individuals participating in an epidemiological cohort study. Application of this framework in a London-dwelling subset of a national English cohort provided a comparison in health effect estimates when assigning both exposure assessment methods to investigate the associations between ambient air pollution exposure and cognitive function. Adverse effects of exposure to NO2, PM2.5 and PM10 associated with cognitive function were found in a longitudinal cohort study with long follow-up and repeated measurements. Associations with cognitive test performance when applying personalised exposure estimates were similar for composite memory scores, but a consistent pattern of slightly more adverse effects with executive function scores were observed. To provide context, the decline in memory and executive function scores observed per IQR increase in long-term residential NO2 exposure was found to be equivalent to that of ageing by approximately 1.5 and 4 years, respectively, in the ELSA cohort [23].

Personalised concentration estimates were found to be markedly lower than residence-based outdoor concentration estimates on average for all pollutants investigated. This is likely due to the fact that Londoners spend up to 95% of their time indoors [10] where exposure to ambient pollutants is lower depending on factors such as the indoor/outdoor ratio of concentrations and infiltration efficiency [13]. A recently published review of studies aiming to personalise estimates of exposure to ambient air pollution also reports that individuals spend the majority of their time indoors at home (60–70%), as well as a large proportion of their time in other indoor environments, although these proportions differ between population groups and individuals [35]. The LHEM has not been validated against personal measurements given the difficulty in designing a personal monitoring campaign of the required size to perform such validation. The present study is unaware of any such study to date. However, CMAQ-urban has been validated against measured concentrations [6] (Supplementary Material Fig. S1 and Table S1). Additionally, the LHEM model utilises several microenvironment models to estimate exposure to ambient sources when indoors in specific building types, as well as when inside different vehicles. Infiltration rates of ambient pollutants into indoor environments differ by both building and vehicle type, as well as by pollutant, which is accounted for by an indoor/outdoor building infiltration model [33] that accounts for 15 Greater London-dwelling archetypes (covering 76% of the known London-dwelling stock) and several independent mass balance equations for each transportation mode. Measurement campaign data conducted across the London Underground network were also included in LHEM modelling of microenvironment concentrations [15].

Previous work investigating personalised exposure in comparison to outdoor residence-based estimates has often not found such marked differences in estimated concentrations between exposure assessment methods [17, 18, 35]. However, infiltration of ambient concentrations to indoor environments is often not taken into account and methodologies employed in apportioning microenvironment exposure (as well as the microenvironments accounted for) currently differ greatly between such studies. In fact, assessing personalised exposure, even if making no significant difference in the effect estimates, reflects a more accurate exposure, in terms of magnitude, to real personal exposure.

Despite the observed reductions in concentrations of personalised exposures to each pollutant in comparison to outdoor residence-based estimates in the present study, correlations between the two were high for NO2 (0.92) and PM2.5 (0.94), but somewhat lower for PM10 (0.68). Previous studies have reported similarly high correlations for NO2 in adult populations [17, 18], as well as for PM2.5 in children [19, 20]. No previous work identified by the present study has reported on the correlation between residence-based and personalised estimates of PM10.

High correlations between the two exposure assessment methods likely explain the similarities observed in effect estimates and such findings may suggest that the inclusion of time-activity information into personalised exposure estimates may not improve epidemiological analyses assigning residence-based estimates only. Furthermore, if the difference between personalised and residence-based estimates is relatively uniform across individuals and does not greatly affect the rank order in exposure per pollutant, then a difference in health effect estimates between the two methods would not necessarily be expected. It should also be noted that the assignment of ambient concentrations in epidemiological studies provides surrogates of exposure and may not necessarily reflect true personal exposure, which is likely lower, for many of the reasons described in the present study regarding time spent in various microenvironments. The personalised concentrations assigned in the present study (which are lower than modelled ambient concentrations) cannot be interpreted as having a “smaller” effect on cognition, as they are a reflection of the surrogate measures, hence effect estimates are presented by IQR and not by a specific change in concentration.

Previous work has generally reported similar findings to that of the present study, with the application of personalised exposure providing little difference in health effect estimates in comparison to that of residence-based estimates for mortality [21, 22] and markers of inflammation [36] in adult populations, as well as lung function [20] and asthma [37] in children. Letellier et al. [38] did find associations in cardio-metabolic markers when assigning personalised exposure in comparison to no associations observed for ambient residence-based estimates. However, the study assigned residence-based estimates using an average value within a 1600 m buffer around the residence, which may not capture the necessary fine-scale variations in outdoor concentrations between participants required for accurate health effect estimation when using residence-based estimates of exposure. The present study utilised fine-scale spatial modelling of ambient concentrations at a 20 × 20 m scale.

Adaptation of the LTDS as an external data source of population time-activity and the microenvironment modelling capabilities of the LHEM presented here provide a number of advantages over previous studies aiming to personalise assigned exposures to ambient air pollution using travel surveys. Shekarizzfard et al. [39] utilised travel survey data encompassing 15,572 trips (from 5945 individuals) in Montreal, Canada, for the year 2008 in order to model exposure trajectories for vehicle drivers/passengers, public transport users and active commuters using an integrated transportation and emissions model linked to a dispersion model. Comparison between daily average outdoor residence-based exposure to NO2 and time-weighted averages including time-activity patterns found 89.6% of individuals to be assigned lower concentrations when using the outdoor residential estimate only. Similarly, Shekarizzfard et al. [40] calculated time-weighted average exposure to black carbon (BC) and ultrafine particles (UFP) using 2011 travel survey data for 1,179,489 trips (341,274 individuals) in Toronto, Canada, finding the median mobility-based exposure to be 11.6% and 63.2% higher than outdoor residence-based exposure for UFP and BC, respectively. In both cases, exposure to ambient concentrations when indoors or in-vehicle was not taken into account, but ambient concentrations assigned at point-time locations without factoring in infiltration to indoor spaces were applied.

As previously described, the LHEM provides a suite of modelling procedures to estimate the infiltration of ambient air pollution into indoor and in-vehicle microenvironments. This provides a further advantage over previous studies that estimated ambient concentrations at several locations such as home/work/school and assigned time-weighted averages [5, 22, 39], which do not account for actual infiltration of ambient pollutants to indoor/in-vehicle environments. Lane et al. [36] did apportion time in separate microenvironments (inside/outside the home, at work, on the highway or other) to investigate the difference in exposure to UFP when accounting for time-activity in 140 residents of Boston, USA, with a mean age (59.1 years) comparable to that of the present study. Lower exposure to UFP was estimated for those spending a greater amount of time away from the home, however, the calculation of personalised exposure did differ from that of the present study. Residence-based estimates (based on proximity to a major highway) were assigned for workers assumed to be highly exposed to traffic-related air pollution (TRAP) when at work, vehicle type information was unknown and 100% infiltration was assigned when in-vehicle, infiltration to the home was inferred from air conditioning usage and urban background average residential concentrations were assigned to other microenvironments. Beckx et al. [41] utilised a model estimating time-activity patterns based on survey data (~10,000 person-day activity-diaries collected between 1997 and 2001), including the estimation of transportation mode for trips, which was extrapolated to a synthetic population with inferred residential information. Residential modelled concentrations were then compared to personalised exposures (constructed from hourly dynamic exposures accounting for time-activity) for the city of Utrecht in the Netherlands. The results in Beckx et al. [41] are reported in terms of person-hours spent above concentration thresholds for PM10 and PM2.5 for April 2005, with marked increases in time spent above such thresholds observed for personalised exposure in comparison to residence-based estimates.

The LTDS survey provides time-activity information for a large number of participants across multiple years and is a continually ongoing study [10]. The ability of the LHEM to model exposure to ambient air pollutants across a range of microenvironments allows for an almost complete modelled picture of exposure to ambient concentrations and does not require time-weighting of ambient concentrations based solely on point-time location. The capabilities of the LHEM to account for infiltration into indoor or in-vehicle microenvironments provides a plausible explanation for the fact that the present study observed a marked reduction in assigned personalised exposures in comparison to outdoor residence-based estimates; a finding that has not been consistently reported in the literature [17, 18, 39].

Additionally, the epidemiological analysis of cognitive function in the present study utilised a cohort providing up to 15 years of follow-up and 5.6 repeated measures on average. Epidemiological studies of cognitive function thus far generally rely on cross-sectional study designs or longitudinal studies with fewer repeated measures and both of these issues have been discussed in the literature as potential weaknesses that must be addressed in future work [4].

The present study utilised external travel survey data to estimate personalised exposure to ambient air pollutants [10] and applied the estimates to an epidemiological analysis within a cohort. Indirectly estimating personal exposure is a crucial issue as the ability to measure it is often not viable, given the cost involved and the inconvenience it would place on individuals over a long enough period of time. Other methodologies have been previously employed to indirectly estimate time-activity patterns and personalise exposure to ambient air pollution, such as agent-based modelling [19, 42], tracking studies [43] and studies collecting information to estimate time spent between the home and work/school address [22]. However, the use of the LTDS in the present study allowed a breakdown of each individual’s survey day at minute-by-minute resolution, given the detailed collection of trip origin/destination, home/work address and transportation mode, as well as hourly modelling of residential concentrations at a fine spatial scale.

Application of the LTDS and LHEM data to personalise outdoor residence-based estimates does however include several limitations. Principally, the application of LTDS survey data and the LHEM in the present study utilised just age and area of residence to personalise exposures, assuming that these two factors have an impact on the difference between residential concentrations and those adjusted for time-activity patterns. This is likely not the case and future work will explore the use of other information available in the LTDS (such as gender and markers of socioeconomic status) to aggregate exposure factors and investigate the impact on the adjustment of residence-based exposure estimates. At present, the LHEM assigns hourly ambient CMAQ-urban estimated residential concentrations for each LTDS participant for the year 2011, whilst the LTDS data set used spanned 2005–2010. This assumes that the spatial variation of outdoor residential concentrations did not markedly change between year of interview and 2011, as well as expecting that representative time-activity and population mobility information derived for these years applies to the years of ELSA follow-up that this period did not cover (2002–2004 and 2011–2017). One further limitation imposed by the necessity for ELSA participant data to remain anonymised was the categorisation of residence-based estimates based on the distribution of assigned postcode estimates. This process may have affected the adjustment of residential modelled estimates by diluting the between-person variability in exposure to some extent (discussed further in Wood et al. [23]).

The LHEM models point-time exposure to ambient air pollutants at minute-by-minute resolution for LTDS individuals when outdoors, indoors and in several transport microenvironments. It does not model exposure to indoor-generated air pollution. Investigation into indoor-generated air pollution in London homes estimated indoor sources to increase indoor concentrations of NO2, PM10 and PM2.5 by 26–37% [44]. The aim of the present study was to assess the potential for utilising the LTDS and LHEM to indirectly adjust modelled ambient residence-based concentrations which are often assigned in epidemiological studies, allowing for comparisons between health effect estimates using a more traditional approach and one aimed at personalising ambient exposure estimation through the inclusion of time-activity and population mobility information. Adaptation of the LHEM to incorporate indoor-generated air pollution in future work would give a more complete picture of total personal exposure but this was not the aim of the present study.

The incorporation of time-activity information to personalise exposure assessment in epidemiological study generally relies on the inclusion of indirect information and this is unlikely to change in the near future. Direct measurement of personal exposure is currently difficult to achieve for the purposes of viable epidemiological analyses. The development of methodologies to accomplish the accurate personalisation of estimates will depend on the availability of data from sources such as detailed time-activity surveys and the application of microenvironment modelling techniques. The present study provides a novel framework and case study comparing the assignment of residence-based and personalised estimates in an epidemiological analysis of cognitive function, finding little difference in health effect estimates. Future work will aim to develop the presented framework and assess the calculation of personal exposure in London-dwelling individuals by factors other than age and area of residence, as well as applying the framework to other health outcomes.