Estimating PM₂.₅–influenza risk using home–school–commute exposure modeling from hourly monitoring data: schoolchildren in Guangzhou, China, 2014–2019

Abstract

Background

Many studies assessing urban air-pollution health impacts assign outdoor PM₂.₅ at a single residential address, ignoring children’s daily movements between home and school. This simplification may misclassify exposure and bias short-term risk estimates, potentially misguiding school-area prevention priorities.

Methods

Using 2,543 laboratory-confirmed influenza cases among school-aged children in Guangzhou (2014–2019) and hourly monitoring–derived PM₂.₅ concentration fields, we tested whether mobility-aware exposure assignment changes short-term PM₂.₅–influenza risk estimates. We constructed a residence-based model (RBM) and a multi-context activity-weighted model (MCAWM) that reallocates hourly PM₂.₅ across home, school, and street-network commute corridors using stylized school-day schedules and open-source street networks.

Results

Although RBM and MCAWM lag-0 daily PM₂.₅ estimates were highly correlated (r = 0.93), 15.8% of child-days differed by ≥ 10 µg/m³. Per 10 µg/m³, cumulative relative risks were 1.04 (95% CI: 1.02–1.06) for RBM and 1.10 (1.06–1.14) for MCAWM, with improved fit and slightly more persistent lag effects under MCAWM. The concentration-dependent difference in cumulative relative risk between MCAWM and RBM increased above ~ 20 µg/m³, with a breakpoint near 45 µg/m³ indicating more rapid RBM underestimation at higher concentrations. Stratification by school environment revealed the largest mobility gains in high-density, low-greenspace schools. Scenario-based population attributable fractions indicated that, at 50 µg/m³, residence-only exposure would miss roughly 0.20 additional attributable cases per 1,000 children per season.

Conclusions

A synthetic “ground-truth” experiment and sensitivity analyses suggest these gains reflect structural reductions in exposure misclassification rather than model tuning. Aligning hourly PM₂.₅ with children’s home–school–commute activity spaces indicates that residence-only assignment tends to attenuate the estimated PM₂.₅–influenza relationship and can understate preventable burden, supporting scalable targeting of school-area interventions such as school greening, near-school traffic management, and lower-exposure commuting corridors.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12889-026-27252-0.

Background

Ambient fine particulate matter (PM₂.₅) is a major environmental determinant of human health, with well-established associations with both respiratory and cardiovascular morbidity and mortality, including from acute respiratory infections such as influenza. Children, and especially school-aged children, are among the most vulnerable populations: they have developing respiratory and immune systems, higher ventilation rates per body weight, and spend substantial time in dense indoor and semi-enclosed settings where exposure and transmission risks can be amplified. Although much of influenza transmission in school settings occurs indoors, ambient outdoor PM₂.₅ remains epidemiologically relevant because it can infiltrate classrooms and other school buildings to varying degrees, while also contributing to children’s exposures during commuting, school arrival and departure periods, and time spent in semi-enclosed school spaces. Numerous studies across diverse settings have reported short-term associations between elevated PM₂.₅ levels and increased incidence or hospitalizations due to influenza-like illness (ILI) or laboratory-confirmed influenza cases [1–3]. Most of these epidemiologic studies rely on ambient outdoor PM₂.₅ metrics, typically derived from monitoring stations or modeled concentration surfaces, rather than direct measurements of indoor concentrations or personal exposure. These findings are biologically plausible. PM₂.₅ has been shown to impair mucociliary clearance, disrupt respiratory epithelial integrity, and increase host susceptibility to viral infections via oxidative stress, inflammation, and immune dysregulation [4]. Moreover, PM₂.₅ may act as a carrier for viruses including influenza, facilitating airborne transmission and increasing viral load in the ambient environment [5–7]. These mechanisms suggest that ambient PM₂.₅ can increase infection risk by contributing to the inhaled particulate burden encountered across both outdoor and indoor microenvironments, even when the resulting infection transmission occurs predominantly indoors.

Despite this strong evidence base, most related urban health studies continue to rely on simplified exposure assessment methods. The prevailing approach estimates PM₂.₅ exposure at an individual’s residential address, typically by mapping that location onto monitor-based concentration fields or modelled surfaces, thereby assuming residential values fully represent personal exposure [8, 9]. While convenient and compatible with address-linked health records, this method implicitly assumes that individuals spend the majority of their time at home, a presumption that often fails to reflect actual time–activity patterns, especially for children. This spatial disconnect between where exposure is measured and where it actually occurs has been termed the “residential exposure fallacy”, rooted in the broader Uncertain Geographic Context Problem (UGCoP) [10]. In reality, children spend large portions of the day at school or commuting between home and school, environments that may exhibit substantially different pollution profiles from their residences. Emerging literature indicates that failing to account for such non-residential exposures may attenuate observed associations or introduce bias in risk estimates [11, 12]. Accordingly, the issue in the present study is not whether indoor transmission occurs, but whether a more behaviorally realistic assignment of ambient outdoor PM₂.₅ across children’s daily activity contexts better captures the exposure patterns relevant to influenza risk than residence-only exposure assignment.

Attempts to reduce this bias have begun to appear. In schoolchildren research, one line of work has estimated traffic-related pollution separately at children’s homes and schools using dispersion models, showing independent effects of each microenvironment on respiratory outcomes, while another has constructed time-weighted exposure metrics by combining high-resolution air-pollution maps with population-average activity patterns from national mobility surveys [13, 14]. In adult populations, studies have applied questionnaire-based or GPS-derived hourly profiles [15, 16] and more recently, mixed-methods mobility proxies [17]. However, these approaches still rely largely on broad temporal averages that smooth over substantial within-day variation, have rarely been implemented in large, multi-year schoolchildren cohorts, and focus predominantly on chronic or non-infectious outcomes. As a result, the influence of fine-scale, mobility-driven PM₂.₅ dose in schoolchildren on acute infectious-disease risk, such as influenza, remains largely untested.

Recent advances in geospatial technologies, high-resolution environmental monitoring, and open-source software now provide new opportunities to address these challenges. These tools have significantly enhanced our ability to model individual-level environmental exposure with high spatial and temporal precision [18–20]. In infectious disease research, these developments, coupled with growing availability of individual-level surveillance data, including geocoded home and school addresses, make it feasible to reconstruct dynamic, multi-location exposures that better reflect real-world behavior. Methodological frameworks such as multi-location time-weighted modeling (e.g., exposure across home, school, and commuting paths) and trajectory-based exposure modeling now support more realistic assessments of spatiotemporal exposure patterns [21–25]. However, these approaches have rarely been applied to short-term influenza outcomes or explicitly tailored to school-aged children, and most existing work treats mobility-informed metrics as descriptive add-ons rather than directly quantifying how they change estimated health risks. Specifically, we still lack a mobility-informed, multi-context PM₂.₅ modeling framework that is structured around school-day timetables and commute corridors, implementable with routine home–school surveillance data, and designed to show when, where, and by how much residence-only exposure understates acute schoolchildren influenza risk and its preventable burden.

To address this gap, we implemented and evaluated a practical home–school–commute exposure assignment approach for influenza among schoolchildren in Guangzhou (2014–2019), comparing two individual-level PM₂.₅ exposure metrics derived from hourly ambient concentration fields and incorporating readily available, stylized school-day mobility information. The first is a conventional residence-based model (RBM) that aggregates hourly ambient PM₂.₅ at each child’s home into daily exposure. The second is a multi-context activity-weighted model (MCAWM) that reallocates hourly exposure across home, school, and street-network commute corridors using fixed school-day schedules and open street-network routing between home and school. Both exposure series were derived from the same monitoring network and meteorological fields and were linked to influenza onset using distributed lag non-linear models (DLNMs). Importantly, our objective is to improve the behavioral and spatial allocation of ambient outdoor PM₂.₅ exposure, rather than to reconstruct indoor concentrations or personal dose directly. We then asked four questions: (1) how RBM and MCAWM differ in estimated PM₂.₅–influenza associations and lag–response structures; (2) whether mobility-informed modeling reduces exposure misclassification and yields larger excess risks at higher pollution levels; (3) in which school environments the gains from MCAWM are greatest; and (4) how these differences translate into population-attributable influenza burden under realistic PM₂.₅ policy thresholds. In addition, we implemented a simulation-based “synthetic truth” experiment to benchmark RBM and MCAWM against known exposure fields and used extensive sensitivity analyses to test whether observed advantages reflect robust, context-consistent improvements in exposure assignment rather than artefacts of particular modeling choices.

The novelty of this study lies in linking mobility-informed exposure reconstruction to epidemiologic risk estimation, rather than treating mobility-informed metrics as descriptive refinements alone. By explicitly comparing residence-only and home–school–commute exposure assignment, we show how a scalable multi-context framework can reduce structured exposure misclassification, improve the spatial realism of ambient PM₂.₅ assessment, and generate more policy-relevant evidence for school-area environmental health interventions.

Materials and methods

A schematic overview of the exposure modeling and epidemiological analysis workflow is shown in Fig. 1. This figure summarizes the main analytical steps of the study. De-identified influenza case records were linked to geocoded home and school locations, together with hourly PM₂.₅ data from the CNEMC monitoring network, daily meteorological variables from ERA5-Land, and OpenStreetMap-derived commuting routes. Two exposure metrics were then constructed: a residence-based model (RBM), based on the home location only, and a multi-context activity-weighted model (MCAWM), which redistributed hourly PM₂.₅ across home, school, and commuting contexts before aggregation to a daily exposure metric. These daily exposures were linked to influenza onset using a time-stratified case–crossover design and distributed lag non-linear models (DLNMs). Downstream analyses included concentration-dependent risk comparison, school-context stratified analyses, scenario-based population attributable fraction estimation, sensitivity analyses, and simulation-based validation against synthetic ground-truth exposure fields.

Fig. 1.

Schematic workflow of exposure modeling and epidemiological analysis

Data sources

Schoolchildren influenza data

We obtained de-identified, individual-level records of schoolchildren influenza cases from the Guangzhou Center for Disease Control and Prevention (CDC), covering the period from May 1, 2014, to December 31, 2019. These cases were reported through the National Disease Surveillance Information Management System (NDSIMS), which compiles daily submissions from 259 sentinel facilities across all 11 administrative districts. Each record included symptom onset date, age, sex, and diagnosis type, classified as either clinician-confirmed or laboratory-confirmed based on national criteria. Laboratory confirmation was performed using RT-PCR, antigen detection, serology, or viral culture. To ensure comparability in daily mobility patterns, we restricted the sample to children approximately aged 6–15 years old who were enrolled in primary or middle school. This school group exhibits relatively consistent weekday routines and commuting behavior, reducing heterogeneity in exposure modeling. The study window was bounded to exclude the COVID-19 pandemic period, which disrupted influenza surveillance and school attendance. After age-based selection, 15,839 cases were initially identified. Residential addresses were geocoded using the Baidu Maps API and transformed from BD-09 to WGS-84 coordinates for compatibility with OpenStreetMap-based routing networks. School locations were geocoded manually using official school registries and guardian-provided information, with manual verification to resolve naming inconsistencies and spatial ambiguities.

We then removed cases with a home–school straight-line distance > 20 km to reduce exposure misclassification from likely heterogeneous motorized commuting that could not be reliably approximated within the stylized MCAWM framework. Cases exceeding this threshold accounted for only a very small proportion of those with valid home–school distance information (1.7%), consistent with the local context of nearby school enrollment among primary and middle school students in China. While this restriction may slightly narrow generalizability to children with more local and regular school-day mobility patterns, its influence on the overall exposure distribution and effect estimates is expected to be limited. We further restricted the analytic sample to weekday-onset cases because the MCAWM timetable explicitly represents regular school-day mobility rather than heterogeneous weekend routines. This restriction was intended to improve alignment between the school-day exposure schedule and the case–crossover index day, rather than to imply that only same-day or weekday exposure is biologically relevant to influenza onset. Because influenza has a short but non-zero incubation period and the DLNM framework incorporated lagged exposures over 0–14 days, weekend PM₂.₅ exposures preceding weekday onset were still captured through the lag structure. After these exclusions, 2,766 school-aged influenza cases had valid spatial and temporal data for multimodal exposure modeling; subsequent exclusions for missing covariates yielded a final analytic sample of 2,543 cases.

Air pollution data

Hourly concentrations of six criteria pollutants (PM₂.₅, PM₁₀, SO₂, NO₂, O₃, and CO) were obtained from 11 national regulatory air-quality monitoring stations in Guangzhou through the China National Environmental Monitoring Centre (CNEMC) real-time data platform. Station coordinates (WGS-84) were used to georeference each site. These stations are part of China’s routine compliance monitoring network and provide continuous, quality-controlled hourly measurements over multiple years, enabling consistent hour-resolved exposure assignment across the full study period.

As shown in Table S1, the 11 stations are distributed across multiple districts of Guangzhou, including several urban sites and designated control/reference sites, providing broad spatial coverage for city-wide exposure modeling. In this study, the CNEMC hourly observations served as the common exposure backbone for all exposure metrics, ensuring that differences between exposure models reflect alternative representations of children’s time–activity patterns and micro-environment transitions rather than differences in the underlying measurement source.

We recognize that high-resolution, full-coverage PM₂.₅ products for China have been developed through multi-source data fusion and geostatistical or machine-learning modeling. However, publicly available products that are consistently accessible at hourly resolution over a multi-year historical period comparable to our study window (2014–2019) remain limited, whereas widely used open products are more commonly provided at daily resolution. Because our exposure framework requires hour-resolved assignment to align with children’s home–school–commute schedules and to preserve temporal comparability across the full study period, we relied on the CNEMC regulatory network as a transparent and time-consistent hourly data source and treated it as the common exposure backbone for both RBM and MCAWM. Hourly PM₂.₅ data were used to support within-day exposure allocation across home, school, and commuting contexts. The resulting hour-specific assignments were then aggregated into daily exposure metrics for the epidemiological analysis. Thus, the use of hourly data served a methodological purpose in exposure reconstruction rather than implying that the health analysis or policy interpretation should be based on the same time scale as regulatory air-quality standards.

We acknowledge that fixed-site monitoring cannot fully capture micro-scale gradients (e.g., near-road increments) at each individual address. To evaluate the robustness of our station-based exposure assignment, we performed hour-level leave-one-station-out cross-validation and conducted interpolation-parameter and station-configuration sensitivity analyses; details are provided in Section "Residence-Based Model (RBM)" and Section "Sensitivity analyses" and Section "Simulation-based validation and sensitivity analyses".

Meteorological data

Daily meteorological variables were obtained from the ERA5-Land hourly reanalysis product provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), accessed via the Google Earth Engine (GEE) platform and restricted to the administrative boundary of Guangzhou. For each day between May 1, 2014, and December 31, 2019, we aggregated hourly ERA5-Land data to compute five daily meteorological indicators. Air temperature and dew point temperature at 2 m above ground were averaged across the day and converted from Kelvin to degrees Celsius. Relative humidity was calculated using the Tetens formula. Daily total precipitation was computed by summing hourly values and converting from meters to millimeters. Wind speed was derived from the vector mean of eastward and northward 10 m wind components, and downward surface solar radiation was summed and converted from joules per square meter to megajoules per square meter. These five variables, including daily mean air temperature (T2M), relative humidity (RH), total precipitation (PRCP), wind speed at 10 m (WS10), and solar radiation (SRAD), were exported as multi-band GeoTIFFs at ~ 9 km (0.09°) resolution, clipped to the Guangzhou boundary. Meteorological values were then extracted based on each participant’s residential coordinates and synchronized with the pollutant exposure estimates. Given the coarse spatial resolution of the meteorological data, the same daily values were applied uniformly across all exposure segments (home, school, and commuting routes).

Exposure assessment

To evaluate how exposure modeling influences estimated health effects, we developed two progressively realistic pollutant exposure metrics: a conventional residence-based model (RBM), and a multi-context activity-weighted model (MCAWM). Both methods used the same underlying monitoring network but varied in how they represent children’s time–activity behavior and micro-environment transitions.

Residence-Based Model (RBM)

The RBM assumes that each child remains at home throughout the 24-hour day. Hourly ambient PM₂.₅ concentrations were interpolated to the home coordinate using inverse distance weighting (IDW, power = 2) based on the four nearest of 11 national monitoring stations. The daily exposure was calculated as the simple mean of the 24 interpolated hourly values:

1

IDW was preferred over kriging or land-use regression for pragmatic and statistical reasons. With 11 monitoring sites, variogram-based kriging would require strong assumptions and may yield unstable spatial covariance estimates, while developing a reliable land-use regression would be underpowered and prone to overfitting, particularly for an hour-resolved framework where temporal variability cannot be captured by static land-use predictors alone without additional time-varying inputs. We therefore adopted a parsimonious station-based interpolator and assessed its performance using hourly leave-one-station-out cross-validation (details below).

To evaluate the robustness of this interpolation scheme, we performed a leave-one-station-out hourly cross-validation for one representative contiguous week per year (2014–2019), exploring a grid of IDW parameters (k = 3–8 nearest stations; power = 1.0–3.0). For PM₂.₅, the best configuration yielded an RMSE of 10.97 µg/m³ and an R² of 0.65 (N = 9,957 station-hours); for ozone (O₃), the corresponding R² was 0.79. These results indicate that IDW provides an adequate approximation of monitored concentrations at unmeasured locations, supporting our use of this method in the RBM exposure assessment. The final RBM implementation used the conventional configuration of k = 4 and power = 2, which showed very similar performance to the cross-validated optimum while keeping the interpolation scheme simple and comparable with previous studies.

Multi-Context Activity-Weighted Model (MCAWM)

MCAWM extends the RBM by redistributing the 24-h exposure across three activity contexts, including home, school, and commute, rather than assuming that children remain at home all day. Based on typical school schedules in Guangzhou, each day was partitioned into fixed hourly blocks: 00:00–06:59 and 17:00–23:59 were assigned to home, 08:00–15:59 to school, while 07:00 and 16:00 were split equally between home and commute (0.5/0.5), yielding a total commute time of 1 h per day. For each child, home and school locations were matched to the nearest nodes on walking, cycling, and driving networks derived from OpenStreetMap. For each mode, we computed the time-shortest path between home and school and resampled three equally spaced points along the route. Commute-mode weights were determined by the shortest walking travel time t_walk: <12 min (walk/bike/drive = 0.70/0.25/0.05), 12–35 min (0.30/0.50/0.20), and > 35 min (0.10/0.30/0.60). At each hour h of day d, ambient PM₂.₅ concentrations at home, school, and along each modal route were obtained from the same IDW surface used for the RBM. The hourly activity-weighted concentration was:

2

Where are the time fractions assigned to each context and are the mode-mix weights for walking, cycling, and driving. The daily MCAWM exposure was then defined as:

3

It represents a time-weighted, multi-context ambient PM₂.₅ exposure that accounts for both school attendance and realistic commuting patterns.

MCAWM is a stylized mobility-informed ambient exposure model rather than an observed trajectory model. It approximates schoolchildren’s typical weekday exposure using routinely available home–school information, open street networks, and plausible time–activity schedules. As such, uncertainty remains in the use of fixed school timetables, inferred commute modes, modeled rather than observed routes, and simplified time allocation across home, school, and commuting contexts. These assumptions may affect absolute exposure estimates for some child-days, but were adopted to provide a transparent and scalable representation of school-day mobility under routine surveillance data constraints.

Simulation-based validation of RBM vs. MCAWM

Because true individual-level PM₂.₅ dose is unobservable in routine surveillance data, we conducted a synthetic “ground-truth” experiment to compare RBM with MCAWM. We generated an idealized 20 km × 20 km urban domain over multiple days at hourly resolution, with a known spatiotemporal PM₂.₅ field, and placed 11 monitoring stations whose time series were obtained by sampling this field. For both RBM and MCAWM, ambient PM₂.₅ at any location and hour was reconstructed from station observations using the same inverse distance weighting (IDW) scheme, consistent with the empirical exposure assessment.

We then simulated ≈ 150 school-age children with randomized home and school locations and assigned walking, cycling, or driving modes based on home–school travel time, following the MCAWM “time-first” logic. For each student-day, the true daily exposure was defined as the 24-hour mean of the true field sampled along the full trajectory (home, school, and straight-line commute). RBM exposure used only the 24-hour home concentration, whereas MCAWM used a 24-hour time-weighted mean across home, school, and commute locations. We compared daily RBM and MCAWM estimates with the true exposure using RMSE, MAE, bias, and the proportion of student-days with smaller absolute error (“rank wins”). Across multiple random seeds, MCAWM consistently reduced error and bias and was closer to the truth on most days, supporting its use as the primary PM₂.₅ exposure metric in the main case–crossover DLNM analyses.

Study design

We adopted a time-stratified case–crossover design to examine the short-term association between ambient PM₂.₅ exposure and laboratory-confirmed schoolchildren influenza in Guangzhou from May 2014 to December 2019. This design is ideal for investigating acute effects of transient exposures, as each case acts as its own control, thereby inherently adjusting for time-invariant confounders including age, sex, and socioeconomic status. For each case day, control (referent) days were selected from the same year, calendar month, and day of the week, thereby adjusting for long-term trends, seasonal variation, and weekly cycles. However, this matching strategy may not fully account for short-term fluctuations in influenza activity within a month, such as localized outbreak peaks, and we therefore interpret the absolute magnitude of the estimated associations with this limitation in mind. Because the main objective of the present study was to compare RBM and MCAWM under the same matched design and case series, such residual temporal fluctuation is less likely to explain their relative differences. Exposure was modeled using two previously described methods (RBM and MCAWM). To assess potential nonlinearity in the exposure–response relationship, we specified distributed lag nonlinear models (DLNMs) with natural splines (df = 2) along both exposure and lag dimensions. As shown in Figure S1, while the fitted curves were flexibly modeled, all exposure metrics displayed approximately linear cumulative risk patterns over the main observed PM₂.₅ range in the analytic data (approximately 0–60 µg/m³). Therefore, we adopted a linear exposure–response function in the final DLNM models to enhance interpretability and reduce model complexity.

The study was approved by the Human Research Ethics Committee of the affiliated university. A total of 2,543 influenza cases were included in the final analytic sample after excluding records with missing values for covariates.

Statistical analysis

We constructed distributed lag non-linear models (DLNMs) for each exposure metric using a linear exposure–response specification and a natural cubic spline (df = 2) for lag (0–14 days). To facilitate direct comparison between exposure metrics, all models were centered at a common reference value, defined as the 5th percentile of the lag 0 RBM PM₂.₅ distribution across all case and referent days. This shared centering facilitates direct comparison of cumulative risk estimates across exposure models, while avoiding extrapolation beyond the observed range for either metric. Models were estimated using conditional logistic regression (clogit) under the case–crossover framework, with strata defined by matched year, month, and weekday. The DLNM model was specified as:

4

Where: is a stratum-specific intercept adjusting for year, month, and weekday; is the cross-basis matrix for PM₂.₅ exposure over lag 0 to 14 days (15-day lag window), assuming a linear exposure–response and natural spline lag-response with 2 degrees of freedom; is the cross-basis for ozone (O₃), modeled identically to PM₂.₅; denotes cross-basis matrices for five meteorological covariates: temperature (T2M), relative humidity (RH), wind speed (WS10), precipitation (PRCP), and solar radiation (SRAD), each with linear exposure–response and natural spline lag-response (df = 2); Weekend and Holiday are binary indicators controlling for day-of-week mobility and healthcare-seeking variation.

Although the time-stratified case–crossover design adjusts for broad temporal trends, explicit adjustment for weekends and national public holidays (130 days, based on official calendars, May 2014 – December 2019) was included to capture residual variation in exposure and healthcare-seeking behavior. Daily exposure and meteorological values were assigned based on the residential location of each case and referent day. Greenness and population density do not vary meaningfully at the daily scale and were therefore not included as time-varying covariates in the main models; instead, they were used to define subgroups for effect modification analyses.

To avoid multicollinearity, we excluded pollutants (e.g., SO₂, CO, PM₁₀) highly correlated with PM₂.₅ (Pearson r > 0.7). O₃ was retained to adjust for potential co-pollutant confounding. All covariates, including meteorological variables and O₃, were modeled using cross-basis functions to flexibly account for concurrent and delayed effects. Following this selection, the remaining covariates exhibited low pairwise correlations, minimizing the risk of multicollinearity in model estimation. Despite including multiple cross-basis terms, variance inflation factors (VIFs) remained within acceptable bounds across all models, with most below 7 and the highest under 12. While one or two covariates exceeded the conventional VIF threshold of 10, this is expected in distributed lag nonlinear models (DLNMs), where collinearity among lagged terms can inflate VIFs without substantially affecting model stability or interpretability. Cross-basis modeling of environmental covariates aligns with best practice in time-series environmental epidemiology, allowing for flexible control of delayed and nonlinear effects without over fitting.

For each exposure metric, we extracted cumulative relative risks (RRs) over lags 0–14 days associated with a 10 µg/m³ increase in PM₂.₅ above the common reference, along with 95% confidence intervals. We also obtained lag-specific RRs and calculated the area under the cumulative lag–response curve AUC_lag as a summary measure of the temporal risk profile. To quantify how the benefit of mobility integration varies across the exposure range, we generated cumulative exposure–response curves over the observed PM₂.₅ range and quantified the concentration-dependent gain in risk from MCAWM relative to RBM as ΔRR(c)=RR_MCAWM(c)−RR_RBM(c), with 95% confidence intervals obtained from the DLNM prediction variance–covariance matrix. Model performance was compared using Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), with lower values indicating better fit; where appropriate, likelihood ratio tests were also used to compare nested models.

Population attributable fractions (PAFs) were calculated for a series of hypothetical PM₂.₅ policy cutoffs. For each cutoff and exposure metric, we computed the empirical fraction of days with lag-0 PM₂.₅ ≥ and the corresponding cumulative ​ from the DLNM, and then evaluated

5

Differences in PAF between MCAWM and RBM were translated into illustrative added attributable influenza cases per 1,000 children per influenza season using an illustrative baseline incidence assumption of 3 cases per 1,000 children per influenza season.

Stratified analyses

To explore potential effect modification in the association between PM₂.₅ and schoolchildren influenza, we conducted stratified DLNM analyses. We focused on school-level context, defined by school population density and school greenness, because these factors directly capture the crowding and environmental conditions in which children spend much of their day. School-level population density and greenness were calculated within 500 m buffers around the school and dichotomized at their sample-specific medians into “low” and “high” categories. For each subgroup level (for example, high school density or low school greenness), we fitted separate case–crossover DLNMs for RBM and MCAWM using the same 0–14 day lag structure and covariate adjustment as in the main models.

Sensitivity analyses

We ran several sensitivity analyses to assess not only statistical robustness but also uncertainty in the key behavioral assumptions embedded in MCAWM. First, we changed the PM₂.₅ lag window from 0 to 14 days (main analysis) to 0–7 and 0–10 days and refitted the DLNMs for both RBM and MCAWM. Second, we compared PM₂.₅-only models with a multi-pollutant specification that additionally included O₃, SO₂, and CO. Third, we applied method-specific checks: for RBM we re-ran the exposure using three alternative IDW kernels (K = 1; K = 4, power = 1.5; K = 6, power = 3), and for MCAWM we used three variants of the commuting and timetable components (IDW_high_locality, MODE_motorized, TIMETABLE_lunch_evening). Fourth, because seasonal variation may be relevant to both biological plausibility and public-health interpretation, we conducted an additional season-stratified sensitivity analysis using the same linearized DLNM framework, comparing the influenza season (December–May) with the other period (June–November). Fifth, to evaluate whether excluding weekend-related records materially affected the findings, we conducted a supplementary robustness analysis in the subset of records with DiagnoseTime falling on weekends, using the same linearized DLNM framework as in the main analysis. For each specification we re-estimated cumulative 0–14 day RRs per 10 µg/m³ increase in PM₂.₅ and compared AIC and BIC. Details are summarized in Supplementary Table S5.

Results

Descriptive comparison of exposure metrics

Daily PM₂.₅ exposure estimates differed meaningfully between RBM and MCAWM (Table 1; Fig. 1). MCAWM assigned higher average exposure (mean 39.20 µg/m³; median 24.42 µg/m³; IQR 21.99–49.61 µg/m³) than RBM (mean 34.26 µg/m³; median 28.99 µg/m³; IQR 18.23–44.22 µg/m³), indicating that incorporating school-day mobility generally increases the estimated ambient dose. Both exposure distributions were right-skewed, as reflected by mean values exceeding medians, but this skewness was more pronounced for MCAWM, suggesting a longer upper tail and a subset of child-days with substantially elevated exposures under the mobility-informed assignment. Despite this level shift, the two series were strongly correlated (Pearson r = 0.93, p < 0.001), reflecting broadly similar temporal patterns. Bland–Altman analysis showed a mean bias of − 4.94 µg/m³ for RBM relative to MCAWM and revealed that about 15.8% of child-days differed by at least 10 µg/m³, underscoring that mobility-informed modeling can substantially change estimated exposure for a non-trivial share of days even when overall correlation is high.

Table 1.

Summary statistics and pairwise comparisons of lag 0 daily PM₂.₅ exposure across models

Pollutant	Model	Mean	Median	Std Dev	25th %ile	75th %ile
PM2.5	RBM	34.26	28.99	21.42	18.23	44.22
PM2.5	MCAWM	39.2	24.42	22.57	21.99	49.61
Model Pair	Mean bias (µg/m³)	95% LoA lower (µg/m³)	95% LoA upper (µg/m³)	Proportion |diff|>10 (%)	Pearson r	p-value
RBM vs. MCAWM	-4.94	-20.94	11.05	15.81	0.933	0.000

Both RBM and MCAWM exposure distributions were right-skewed, with the skewness more pronounced for MCAWM

Spatial patterns also diverged (Fig. 2). RBM exhibited stronger global spatial clustering (Moran’s I = 0.155, p < 0.001), with high–high local clusters concentrated in central urban districts. MCAWM showed weaker spatial dependence (Moran’s I = 0.06, p < 0.001) and fewer significant clusters, reflecting the smoothing effect of integrating movements between home and school. Difference maps indicated that MCAWM tended to yield lower PM₂.₅ estimates than RBM in high-traffic, school-dense neighborhoods, while differences were smaller or near zero in peripheral areas. These descriptive results suggest that accounting for daily mobility not only changes the average assigned exposure but also reshapes its spatial distribution, particularly in dense inner-city zones.

Fig. 2.

Daily PM₂.₅ exposure estimates and spatial differences across models (lag 0). Notes: Top row: Local Indicators of Spatial Association (LISA) clusters for each exposure model (RBM and MCAWM). Bottom row: spatial differences in estimated PM₂.₅ between the two models, with blue indicating lower estimates and red indicating higher estimates for MCAWM relative to RBM

Primary PM₂.₅–influenza associations and lag–response structures

Cumulative exposure–response curves derived from the residence-based model (RBM) and the multi-context activity-weighted model (MCAWM) are shown in Fig. 3, with corresponding point estimates summarized in Table 2. Both exposure metrics reveal a monotonic increase in schoolchildren influenza risk over the 0–14-day lag window as PM₂.₅ concentrations rise from low to high levels, but the magnitude of association differs substantially between models. Using a unified low-exposure reference (RBM lag-0 5th percentile, 11.5 µg/m³), the cumulative relative risk (RR) per 10 µg/m³ increase in PM₂.₅ was 1.04 (95% CI: 1.02–1.06) under RBM and 1.10 (95% CI: 1.06–1.14) under MCAWM. Model fit indices were also consistently better for MCAWM (AIC = 650,113.6; BIC = 650,242.1) than for RBM (AIC = 650,170.6; BIC = 650,299.1), despite identical degrees of freedom. In addition, visual inspection of the exposure–response curves (Supplementary Figure S1) suggests that MCAWM better captures subtle departures from linearity in the cumulative risk profile at higher concentrations.

Fig. 3.

Cumulative exposure–response curves (lag 0–14) for two PM₂.₅ models. Notes: Shaded areas denote 95% confidence intervals. Although the exposure–response relationship was specified as linear, the cumulative RR curves appear nonlinear due to the summation of lagged effects across the 15-day window

Table 2.

Cumulative relative risk (RR) per + 10 µg/m³ PM₂.₅ increase and model fit indices (lag 0–14)

Metric	Cumulative RR	95% CI (Lower, Upper)	AIC	BIC
RBM	1.04	[1.02, 1.06]	650170.6	650299.1
MCAWM	1.10	[1.06, 1.14]	650113.6	650242.1

All estimates derived from linearized DLNM models controlling for meteorological covariates and calendar effects. RR values represent the cumulative excess risk associated with a + 10 µg/m³ increase in PM₂.₅ above the unified low-exposure reference

Lag–response dynamics were broadly similar between the two models, with both showing elevated risk concentrated in the first two weeks following exposure, but MCAWM suggested slightly more persistent effects. Based on cumulative lag–response functions for a 10 µg/m³ increase in PM₂.₅, the area under the curve (AUCₗₐ_g) was 14.04 for RBM and 14.10 for MCAWM, and the number of lags with statistically significant cumulative RR (95% CI lower bound > 1) increased from 10 days under RBM to 13 days under MCAWM (Supplementary Table S2, Figure S2). These results indicate that replacing residence-only exposure with a simple home–school–commute time-weighted scheme not only strengthens the estimated PM₂.₅–influenza association but also modestly extends the window of elevated risk captured by the distributed lag model.

Concentration-dependent changes in excess risk

Figure 4 and Supplementary Table S3 summarize how the incremental benefit of MCAWM varies across the PM₂.₅ range by plotting the concentration-dependent difference in cumulative risk (ΔRR, lags 0–14) between MCAWM and RBM. Across the main observed PM₂.₅ range in the analytic data, MCAWM and RBM were similar at lower concentrations but diverged progressively as concentrations increased. At very low concentrations, MCAWM produced slightly lower cumulative risks than RBM, and the two models were essentially indistinguishable around 10 µg/m³ (ΔRR near zero). As concentrations rose above roughly 20 µg/m³, MCAWM began to outperform RBM, with ΔRR becoming positive and increasing with concentration. For example, MCAWM yielded a 0.06 higher cumulative RR at 20 µg/m³ and a 0.24 higher RR at 40 µg/m³, with confidence intervals excluding zero at both levels.

Fig. 4.

Concentration-dependent gain in cumulative PM₂.₅ risk when replacing RBM with MCAWM. Notes: a. Difference in cumulative PM₂.₅ risk (ΔRR, lag 0–14 days) between MCAWM and RBM. b. The solid blue curve shows the estimated ΔRR across the observed PM₂.₅ range in the analytic data; the shaded band denotes the 95% confidence interval; the horizontal dashed line marks the null (ΔRR = 0). c. The black line represents the fitted segmented regression, and the vertical dashed line indicates the estimated breakpoint at approximately 45 µg/m³

Segmented regression identified a breakpoint at approximately 44.9 µg/m³, within the observed concentration range. Below this breakpoint, ΔRR increased slowly with concentration, whereas above it the slope nearly doubled, indicating that residence-based misclassification grows fastest during moderate-to-high pollution episodes. This pattern implies that ignoring children’s daily mobility leads to the greatest underestimation of influenza risk precisely on days when PM₂.₅ levels are most elevated, suggesting that residence-based exposure assessments may systematically downplay the health burden of severe pollution events.

Context-specific gains by school environments

To test whether the advantages of the mobility-informed model are concentrated in particular school settings, we stratified children by school-level population density and greenness and re-estimated cumulative PM₂.₅ risks for RBM and MCAWM (Table 3). In high-density schools, MCAWM produced a substantially larger cumulative RR than RBM: 1.17 (95% CI: 1.10–1.23) versus 1.06 (95% CI: 1.04–1.09). In low-density schools, the two models yielded nearly identical estimates (RR ≈ 1.03), with overlapping confidence intervals. A similar pattern emerged for greenness. In low-greenspace schools, MCAWM amplified the PM₂.₅–influenza association relative to RBM (RR 1.08 vs. 1.04), whereas in high-greenspace schools the two metrics produced almost identical cumulative RRs near 1.08.

Table 3.

Cumulative relative risk (RR) per + 10 µg/m³ PM₂.₅ increase (lag 0–14) by school environment and exposure model

Subgroup (school)	Method	RR (95% CI)	Mobility gain†
HighPopSchool	RBM	1.06 (1.04–1.09)	—
	MCAWM	1.17 (1.10–1.23)	+ 10%
LowPopSchool	RBM	1.03 (1.00–1.06)	—
	MCAWM	1.03 (0.97–1.09)	~ 0%
LowGreensSchool	RBM	1.04 (1.01–1.06)	—
	MCAWM	1.08 (1.02–1.15)	+ 3.9%
HighGreensSchool	RBM	1.08 (1.04–1.11)	—
	MCAWM	1.08 (1.01–1.15)	~ 0%

† Mobility gain is defined as (RR_MCAWM/RR_RBM) − 1, expressed as a percentage and based on point estimates. All RRs are cumulative over lag 0–14 days for a + 10 µg/m³ increase in PM₂.₅ above the unified low-exposure reference (RBM lag-0 5th percentile, 11.5 µg/m³), adjusting for O₃, meteorological covariates, weekends, and national holidays

These results indicate that residence-based exposure tends to underestimate PM₂.₅-related influenza risk specifically in high-density, low-greenspace school environments, where children’s exposures during school hours and commuting are most likely to exceed those at home, while the incremental benefit of mobility integration is small in less crowded or greener school contexts.

Public-health impact of mobility-informed exposure modeling

We next translated the differences between RBM and MCAWM into public-health terms by estimating population attributable fractions (PAFs) across a range of hypothetical PM₂.₅ policy cutoffs (Fig. 5; Table 4). At very low hypothetical cutoffs, the PAF curves include negative values. These should not be interpreted as evidence that low PM₂.₅ is biologically protective. Rather, because the DLNM predictions were centered at a shared low-exposure reference of 11.5 µg/m³, cutoffs below or near this reference can yield cumulative RR values below 1 by construction, which in turn produces mathematically negative PAF values. In this setting, substantive interpretation is most appropriate over the higher cutoff range where RR exceeds 1 and the attributable-fraction framework has its usual public-health meaning. Scenario curves showed that mobility integration meaningfully increases the estimated preventable burden of schoolchildren influenza across realistic PM₂.₅ thresholds. As the cutoff rose from 15 to about 40–50 µg/m³, PAFs increased and then gradually declined, reflecting the usual trade-off between higher individual-level risk at extreme concentrations and fewer days above the cutoff. Across almost the entire range, MCAWM yielded consistently larger PAFs than RBM, indicating that residence-based exposure underestimates the share of influenza cases attributable to ambient PM₂.₅. For illustration, at a scenario cutoff of 50 µg/m³, RBM classified 22.8% of child-days as exposed and implied a PAF of 3.7% (95% CI: 2.2–5.0%), whereas MCAWM classified 29.8% of days as exposed and yielded a PAF of 10.4% (95% CI: 7.1–13.2%). This 6.7% point difference corresponds to roughly 0.20 additional attributable cases per 1,000 children per influenza season that would be missed if only residence-based exposure were used.

Fig. 5.

Scenario PAF curves by PM₂.₅ policy cutoff (RBM vs. MCAWM). Notes: Population attributable fractions (PAFs) for schoolchildren influenza across hypothetical PM₂.₅ policy cutoffs, comparing RBM with MCAWM. PAFs were computed as described in the Methods, using cumulative relative risks over lags 0–14 days from DLNMs with a unified low-exposure reference (RBM lag-0 5th percentile, 11.5 µg/m³). Shaded ribbons denote 95% confidence intervals. Negative PAF values at very low cutoffs reflect the centering of RR predictions at 11.5 µg/m³; for cutoffs below or near this reference, predicted RR may fall below 1, yielding mathematically negative PAF values. These values should not be interpreted as biologically protective effects

Table 4.

Scenario population attributable fractions (PAFs) for schoolchildren influenza at a PM₂.₅ cutoff of 50 µg/m³, by exposure model

Cutoff (µg/m³)	Method	Exposed share (%)	Cumulative RR	PAF % (95% CI)	ΔPAF vs. RBM (%)	Added attributable cases per 1,000 children per season*
50	RBM	22.8	1.191	3.7 (2.2–5.0)	—	—
50	MCAWM	29.8	1.537	10.4 (7.1–13.2)	6.7	0.20 (0.06–0.33)

a. The cumulative RR shown in this table is the model-predicted cumulative RR at the policy cutoff of 50 µg/m³ relative to the shared reference (11.5 µg/m³), used for the PAF calculation. It is therefore not directly comparable to the per +10 µg/m³ cumulative RR reported in Table 2

b. * Added cases were converted from ΔPAF using an illustrative baseline incidence assumption of 3 cases per 1,000 children per influenza season

Simulation-based validation and sensitivity analyses

Our simulation-based experiment provided an external check on these empirical findings. In the synthetic urban domain, MCAWM consistently reduced RMSE and bias relative to RBM and more often produced the closest estimate to the known daily exposure across student-days (Supplementary Table S4), supporting the interpretation that the mobility-informed metric structurally reduces exposure error. Sensitivity analyses further indicated that the main conclusions were robust to alternative lag windows, pollutant adjustment strategies, spatial interpolation kernels, commuting assumptions, seasonal stratification, and weekend-related case restriction (Supplementary Table S5). Shortening the lag window attenuated cumulative RR estimates for both models, although MCAWM retained better model fit than RBM across all lag-window specifications. The additional specification checks likewise supported the overall robustness of the main findings. Especially, in the season-stratified analysis, the PM₂.₅–influenza association was more evident during the influenza season (December–May), with cumulative RR estimates of 1.04 (95% CI: 1.02–1.06) for RBM and 1.10 (95% CI: 1.06–1.14) for MCAWM, whereas estimates during the other period were weaker and imprecise, with confidence intervals including the null for both models. These results suggest that the observed PM₂.₅–influenza relationship is seasonally concentrated and most clearly detectable during the period of higher influenza circulation. In the weekend sensitivity analysis, associations for both models were weaker and imprecise, suggesting attenuation rather than inflation of the positive association observed in the primary weekday-based analysis.

Discussion

This study shows that a pragmatic mobility-informed exposure assignment strategy, implemented using hourly monitored PM₂.₅ concentrations and contrasting a conventional residence-based model (RBM) with a multi-context activity-weighted model (MCAWM), can materially change estimates of short-term influenza risk. By reallocating daily exposure across children’s home, school, and street-network commute corridors, MCAWM consistently yielded stronger and more persistent associations than RBM across complementary lenses, including overall exposure–response strength, lag–response structure, concentration-dependent risk, and school-context stratifications. The largest discrepancies occurred at higher PM₂.₅ levels and in high-density, low-greenness school environments, indicating that exposure error is not random but systematically shaped by school-day mobility and urban form. When translated into scenario-based population attributable fractions, these differences imply that residence-only exposure can miss a non-trivial share of influenza burden on polluted days, precisely in the school settings where children spend substantial daytime hours and commute through traffic-influenced microenvironments. A simulation-based “synthetic truth” experiment and extensive sensitivity analyses further suggest that these gains reflect reduced structured exposure error under common data constraints rather than artefacts of particular modeling choices, supporting the value of mobility-informed exposure assignment for school-area decision support. These findings indicate that the contribution of the present study is not only methodological but also translational: mobility-informed exposure assignment improves the spatial relevance of exposure assessment, reduces structured underestimation of PM₂.₅-related influenza risk, and better supports the prioritization of interventions in high-risk school environments.

Consistent with prior urban health studies [26–29], we also observed that higher daily PM₂.₅ levels were associated with increased risk of schoolchildren influenza. Our results extend this evidence by showing that the magnitude and persistence of this association depend strongly on how exposure is defined: residential metrics, although convenient and often highly correlated with mobility-informed measures, appear to understate acute risk in schoolchildren whose daily routines are tightly structured by school attendance and commuting patterns [17]. This pattern reflects a broader “uncertain geographic context” problem [10] and parallels findings for green-space exposure, where mobility-informed metrics reduce residential underestimation of park use and related health outcomes [26], which our study now extends to short-term air-pollution epidemiology in schoolchildren. Compared with recent meta-analytic estimates that often report relatively small percentage increases in respiratory outcomes per 10 µg/m³ of PM₂.₅ [30], our cumulative RRs fall toward the upper end of the reported range but remain plausible given the multi-day lag window, the particular susceptibility of school-aged children, and the sharper exposure contrasts generated by MCAWM. MCAWM also slightly extended the window of elevated risk and improved model fit relative to RBM, consistent with reduced temporal misclassification and partial recovery of underlying lag–response signals predicted by classical nondifferential measurement error theory. This improvement in fit likely reflects better temporal–spatial alignment between assigned exposure and the biologically relevant exposure window. Whereas RBM assigns all daily PM₂.₅ to the home location, MCAWM reallocates hourly exposure across home, school, and commuting contexts according to school-day mobility, thereby reducing mismatch between when and where exposure is assigned and when and where it likely occurred. In a DLNM framework, such improved alignment can sharpen recovery of lag-specific patterns. More broadly, in a DLNM framework, exposure misclassification may distort not only overall effect magnitude but also the shape of the lag–response relationship itself: true lag-specific peaks may be smoothed, shifted, or diluted across adjacent days, making the temporal pattern appear flatter or less persistent than it actually is. In this sense, the more persistent lag profile observed under MCAWM may reflect not only stronger effect estimation, but also better recovery of the underlying temporal structure of PM₂.₅-related influenza risk.

Our second objective was to understand how residence-based attenuation varies across the PM₂.₅ range. Viewed through a measurement error framework, the contrast between RBM and MCAWM is most consistent with attenuation bias under residence-only exposure assignment. The segmented ΔRR curve indicates that misclassification is modest at low-to-moderate concentrations but grows rapidly once pollution exceeds roughly 45 µg/m³, so contrasts between MCAWM and RBM are greatest during moderate-to-high episodes where clinical burden and policy attention tend to concentrate. Importantly, this exposure misclassification is multidimensional: it includes spatial misalignment, because RBM assigns all exposure to the home location; temporal misalignment, because it cannot align hourly exposure with children’s school-day activity windows; and contextual misalignment, because it ignores differences across home, school, and commuting microenvironments. Mechanistic evidence suggests that PM₂.₅ induces respiratory inflammation, impairs mucociliary clearance, and modulates host immune responses in a dose-dependent manner [31], implying that additional increments of exposure on already polluted days are particularly consequential for infection risk. Consequently, if home-only models systematically miss the extra exposure children receive in schoolyards, playgrounds, and along busy routes precisely on high-pollution days, they will tend to understate both the true health impact of those days and the potential benefits of pollution control.

Our third objective was to identify the contexts in which mobility integration yields the greatest gains. We found that the largest differences between RBM and MCAWM occurred in high-density, low-greenness school environments, whereas estimates converged in low-density or greener school settings. This pattern is consistent with evidence that densely built, traffic-intense corridors exhibit higher and more spatially heterogeneous pollutant levels than greener, less urbanized areas [32]. In such high-density, low-greenness neighborhoods, schools are often embedded in busy road networks with limited vegetative buffering, so children’s exposures during school hours and commuting can systematically exceed those at home. A model anchored solely at the home address will therefore miss much of this “away-from-home” exposure, while MCAWM, by reallocating time to school and commute microenvironments, captures more of this contrast. In measurement-error terms, this misclassification is unlikely to be differential by outcome status, because exposure assignment does not depend on case status, but its magnitude can vary systematically across contextual factors such as school density, greenness, and traffic intensity. High-density schools may also sustain greater baseline influenza transmission because of increased contact and crowding. The larger MCAWM–RBM contrast in these settings therefore likely reflects both reduced exposure misclassification and a transmission context in which PM₂.₅-related effects are more readily expressed. In greener or less crowded school environments, traffic intensity and near-road gradients are weaker, home and school exposures are more similar, and the incremental benefits of mobility integration naturally diminish.

Our fourth objective was to translate these modeling differences into public health terms. Scenario-based PAF curves showed that, under realistic PM₂.₅ policy thresholds, mobility integration consistently enlarged the fraction of schoolchildren influenza attributable to ambient PM₂.₅, especially around commonly discussed standards such as 50 µg/m³. Although the additional burden identified by MCAWM corresponds to only a small increase in attributable cases per 1,000 children per season, it scales to tens or even hundreds of additional cases in a large urban school system over multiple seasons. Moreover, these extra cases are not randomly distributed but are concentrated in high density, low greenness schools in polluted neighborhoods, where children’s exposures and infection risks are already elevated. This pattern suggests that residence only exposure models may understate both the true health burden and the potential benefits of targeted school-based air quality and mobility interventions in precisely those settings where action is most urgent.

Policy-wise, our findings underscore the importance of incorporating behaviorally informed, activity-weighted exposure models into public health surveillance and air quality management. These patterns were robust to multiple sensitivity analyses and to a simulation-based “synthetic truth” experiment, which together suggest that the advantages of MCAWM are not artefacts of any single modeling choice. Interventions for school-aged populations should consider not only residential air quality but also exposures encountered along commute routes and within school environments, where we observed the largest discrepancies between mobility-informed and residence-based models. Scenario PAFs suggest that relying solely on home-based exposure could underestimate both the burden of influenza attributable to PM₂.₅ and the benefits of mitigation, particularly in high-density, low-greenness school contexts. Policies that prioritize children attending schools in densely populated or low-greenness areas may yield disproportionate health gains, especially when combined with green infrastructure and environmental buffers around high-risk school environments during influenza season.

Methodologically, our work shows that substantial improvements can be achieved with a relatively simple mobility-informed exposure model. MCAWM does not rely on personal monitors, GPS traces, or detailed indoor infiltration models. Instead, it leverages routinely available home and school locations, an open street network, plausible commuting routes, and stylized time–activity schedules to construct a time-weighted ambient exposure metric. This places it between crude residence-only approaches and fully individualized dose models: it is substantially more realistic than the former but far more feasible at scale than the latter. In contexts where high-resolution behavioral or microenvironmental data are unavailable, such multi-context, time-weighted ambient models offer a pragmatic pathway to reduce exposure misclassification and sharpen risk estimates for short-term health outcomes. At the same time, MCAWM should be interpreted as a mobility-informed ambient exposure model rather than a direct estimate of personal exposure or inhaled dose. True exposure during children’s daily mobility may still vary with transport mode, vehicle enclosure, ventilation, proximity to traffic, route-side shielding, and individual behaviors such as window opening or trip timing. These factors introduce additional person-level variability beyond what can be captured by modeled ambient concentrations alone, but do not negate the value of MCAWM as a more realistic representation of daily exposure structure than residence-only assignment. This also has practical implications, suggesting that lower-traffic routes, greater roadside buffering, and improved ventilation in travel microenvironments may help reduce exposure during the school commute.

Several limitations should be kept in mind. First, exposure surfaces were interpolated from a limited number of fixed-site monitors and therefore cannot fully capture fine-scale spatial variability, particularly near heavily trafficked streets or in complex street canyons. Second, children’s time–location patterns were based on stylized school-day schedules and inferred commuting behaviors rather than observed GPS or diary data. Accordingly, uncertainty remains in activity timing, time allocation across home, school, and commute, mode assumptions, and modeled routes. These factors may affect child-day-specific exposure estimates, but are unlikely to overturn the overall pattern of stronger associations under MCAWM, which was robust across alternative timetable and commuting scenarios and in the simulation-based validation. Third, we estimated time-weighted ambient outdoor PM₂.₅ rather than personal exposure or inhaled dose. Beyond indoor infiltration, ventilation, window-opening behavior, air-conditioning use, and activity-dependent breathing rates, actual commuting exposure may also vary with transport mode, vehicle enclosure, proximity to traffic, and other microenvironmental conditions not captured by modeled ambient concentrations alone. However, because both RBM and MCAWM were derived from the same outdoor monitoring backbone, these factors are more likely to affect absolute exposure accuracy than the overall pattern of stronger associations under MCAWM. Fourth, although we adjusted for meteorology, weekends, and holidays, unmeasured confounders such as vaccination coverage, household socioeconomic status, or public-health interventions may still influence the observed associations. Fifth, while the time-stratified case–crossover design controlled for long-term trends, seasonality, and weekday-related temporal structure, it may not have fully captured short-term within-month fluctuations in influenza activity, such as localized outbreak peaks. Such residual epidemic dynamics could affect the absolute magnitude of the estimated associations, but are less likely to explain the relative contrast between RBM and MCAWM, which were evaluated under the same matched design and analytic sample. Sixth, we restricted the analysis to weekday-onset cases because MCAWM was defined for regular school-day routines rather than heterogeneous weekend schedules. This was a methodological alignment choice rather than an assumption that only weekday exposure is biologically relevant; weekend PM₂.₅ still contributed through the 0–14 day lag structure. Nonetheless, this restriction may have altered the observed case series and modestly narrowed generalizability. Finally, the study was conducted in a single subtropical megacity with specific school structures, pollution sources, and healthcare-seeking behaviors, which may limit generalizability.

Future research can build on these foundations in two main directions. First, combining ambient-based models with short-term personal monitoring or mobility traces in subsamples would allow calibration of time–activity assumptions and better characterization of indoor–outdoor gradients in schools and homes. Such work could also quantify how commuting mode, route choice, and travel microenvironments shape children’s true exposure during school travel and thereby inform more individualized protective strategies. Second, applying similar modeling strategies to other acute outcomes and to cities with different pollution profiles and school structures would further test the portability of our conclusions and inform local air-quality and school-siting policies.

Conclusion

This study shows that a mobility-informed multi-context activity-weighted model (MCAWM), operationalized as a home–school–commute exposure assignment strategy that reallocates hourly monitored PM₂.₅ across children’s home, school, and street-network commute corridors, can materially alter estimates of influenza risk among schoolchildren, particularly at higher pollution levels and in high-density, low-greenspace school environments. Beyond demonstrating stronger PM₂.₅–influenza associations under MCAWM, this study shows that mobility-informed exposure modeling can meaningfully reduce exposure misclassification, improve the spatial relevance of ambient exposure assessment, and provide a more actionable basis for school-centered environmental health planning. Compared with a conventional residence-based model, the mobility-informed MCAWM strengthened exposure–response associations, revealed greater underestimation of risk during severe pollution episodes, identified specific school contexts where residential metrics perform worst, and expanded the portion of influenza burden that appears attributable and potentially preventable under realistic air quality policies. Because MCAWM is built from data sources already available in many cities, adopting similar mobility-informed exposure assignment can provide an immediate, practical improvement for school-area decision support, helping prioritize interventions such as school greening, traffic management near schools, and lower-exposure commuting corridors.

Authors’ contributions

FJ and JH take responsibility for the integrity of the data and the accuracy of the analysis in this study. FJ, SZ, WM, and YL devised the study concept and design. ZK, GH, and JXH developed the home–school–commute exposure modeling framework and prepared geospatial inputs. TL, MZ, LL, and MM curated the surveillance and monitoring data and contributed to variable construction. ZL, SuZ, and YL contributed to statistical analysis and interpretation. FJ and SZ drafted the manuscript. All authors critically revised the manuscript. SZ, YL, and WM supervised this study and obtained funding.

Funding

This work was supported by the National Natural Science Foundation of China (Grants Nos. 42475183 and 42501270), the 75th Batch of China Postdoctoral Science Foundation General Grant (2024M751122), the National Key Research and Development Program of China (2026ZD01909500), and the Key Project of Medicine Discipline of Guangzhou (No.2025–2027-11).

Data availability

Hourly concentrations of six criteria pollutants (PM₂.₅, PM₁₀, SO₂, NO₂, O₃, and CO), as well as the Air Quality Index (AQI), are available at (https://www.resdc.cn/data.aspx?DATAID=186) .Hourly meteorological variables were obtained from the ERA5-Land hourly reanalysis product provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), accessible via the Google Earth Engine platform at the following link: (https://developers.google.com/earth-engine/datasets/catalog/ECMWF_ERA5_LAND_HOURLY) .The flu dataset used and/or analyzed during the current study are available from the corresponding author on reasonable request. The full analysis code, including implementation of the residence-based (RBM) and multi-context activity-weighted (MCAWM) exposure metrics, the three MCAWM sensitivity scenarios (IDW_high_locality, MODE_motorized, TIMETABLE_lunch_evening), the IDW leave-one-station-out cross-validation, and the synthetic-truth validation experiments, is available at the following link: (https://colab.research.google.com/drive/1NhBhl8eC1gwKsKlouBne55hFrAokNlXw?usp=sharing).

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of Jinan University. The study was conducted in accordance with the Declaration of Helsinki. Informed consent has been obtained from all study participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.