Models that link urban air pollution to people's health over years are built from three different things. The first is a record of what each person breathed, and it depends on the city's sensors. The second is a record of when each person was measured, and it depends on the health system. The third is how the body responds to dose, which is the only part we hope is shared across places. Current practice fits these three together as one model and then either locks it for use in a new cohort or refits it from scratch. When a locked model fails in a new place, nobody can tell which of the three parts broke. This proposal builds the three parts as separate modules. It trains them on a derivation cohort, then moves only the dose-response module to a new cohort, while the sensor-specific exposure module and the health-system-specific visit module are re-estimated locally.
In two sentences, the idea is this. Build a longitudinal pollution-health model whose exposure, visit-timing, and dose-response components can be swapped independently. Then test whether keeping only the dose-response fixed transports better, and gives more consistent effect estimates across cohorts, than either locking or refitting the whole model. The neural continuous-time machinery that dominates the irregular-time-series literature enters only as one candidate backbone. If a plain mixed model with lagged exposure does as well inside the same modular design, we will report that, and it would mean the contribution is the decomposition, not the network.
Core Claim
In pollution-health trajectory models, most of the performance lost when moving to a new cohort comes from two sources that have little to do with biology. The first is how exposure was measured, for example taxi-mounted low-cost sensors, fixed stations, or satellite grids. The second is why and when people were measured, for example annual screening versus visits triggered by illness. A model that keeps these in separate modules and fixes only the dose-response link should close most of the gap between internal and external performance. It should also recover pollution effect estimates that agree across cohorts more closely than those from a single combined model. The evidence that such gaps are large comes from clinical prediction. The OHDSI hemorrhagic-transformation model is a Lasso logistic regression trained on 621,000 US patients with an internal AUC of 0.75. Locked and applied to 10 external databases, it averaged an external AUC of 0.71 and dropped to 0.50 on a German primary-care registry. The authors traced that failure to how data were recorded, not to patient biology. The environmental health literature has almost no equivalent multi-cohort test. Its models are typically checked by leaving out monitors within a single city.
What exists and what is missing
The pieces exist separately, and each paper below plays a distinct role.
Two papers show the gap on the exposure side. The Taiwan air pollution and exercise cohort followed 384,130 adults across 842,394 screening exams. It assigned PM$_{2.5}$ from 1 km satellite grids at home addresses and fit Cox regression with time-varying covariates, finding HR 1.15 (95% CI 1.10–1.20) for high PM$_{2.5}$. It never tested the model outside Taiwan and ignored exposure outside the home. The Xi'an taxi-fleet study did the opposite. It built street-level PM$_{2.5}$ maps from about 200 taxi-mounted sensors over two weeks and computed student commute doses, but it has no health outcomes. The context-aware e-bike study shows why home-address exposure is not the same as dose: cyclists inhale 400–900% more particulate matter than car passengers on the same route. That multiplier enters our exposure module directly.
Three papers supply the method on the modeling side. Mixed Effects Neural ODE (ME-NODE) splits a learned continuous-time vector field into population-level terms (fixed effects) and person-specific terms (random effects), and has been tested only on internal ADNI splits and image benchmarks. Neural Controlled Differential Equations (Neural CDEs) are models whose hidden state is driven by a continuous path built from irregular observations. They reached AUC 0.880 on PhysioNet sepsis and add an observation-count channel so the model can use how often data arrive. Random CDEs reduce training cost by freezing random CDE layers and training only a linear readout.
The adiposity-trajectory study both constrains and motivates the design. It fit classical mixed models to EHR records from about 740,000 people and validated them externally in MVP and the Estonian Biobank. It showed that baseline level and rate of change have largely distinct genetic architectures, with a genetic correlation of 45.2%. It also flagged that sicker people are measured more often, which biases the person-specific estimates.
None of these papers treats the visit process and the exposure measurement process as separate, cohort-specific components. The ME-NODE and Neural CDE work treats observation timing as a nuisance or as a single count channel. The epidemiology work treats exposure as a fixed covariate with no measurement model at all.
Research Program
For person $i$, let $z_i(t)$ be a latent health state observed noisily at visit times $t_{i1}, t_{i2}, \dots$. The model has three parts:
$$ dz_i(t) = f_\theta\big(z_i(t), b_i\big)\, dD_i(t), \qquad D_i(t) = g_c\big(\text{sensor data}_c,\ \text{location and activity}_i\big), \qquad \lambda_i(t) = h_s\big(z_i(t), t\big). $$
The first equation is the dose-response module. Here $f_\theta$ is the shared response to cumulative inhaled dose, and $b_i$ is person $i$'s random effect. The second equation is the exposure module. Here $D_i(t)$ is person $i$'s cumulative inhaled dose, reconstructed by a city-specific function $g_c$ from that city's sensors, with ventilation multipliers for walking and cycling. The third equation is the visit module. Here $\lambda_i(t)$ is the rate at which person $i$ is measured, which depends on their latent state through a health-system-specific function $h_s$. This makes "sicker people are seen more" an explicit part of the model rather than hidden bias.
The two kinds of missing data are handled separately. Gaps in sensor coverage, such as a taxi off duty or a dead battery, enter $g_c$ as uncertainty about dose and carry no information about health. Gaps between visits enter $h_s$ and do carry information about health.
The model is trained jointly on the derivation cohort by maximizing the likelihood of both the observed values and the visit times. We will build it with two interchangeable forms of $f_\theta$. The first is a mixed model with a distributed-lag dose term, which is linear with a spline for time. The second is a mixed-effects Neural CDE driven by the dose path. The Neural CDE uses the random-feature variant as a low-cost fallback if full training does not scale to hundreds of thousands of people.
Transport to a new cohort uses one of three policies:
- Full lock: the whole model is applied unchanged.
- Full refit: every module is re-estimated on target data, with a varying number of labeled target people.
- Modular transport: $\theta$ stays fixed. The exposure module is re-estimated from the target's own sensor data, which needs no health outcomes. The visit module and the random-effect distribution are re-estimated from target visit times and early visits only.
Baselines
The naive reference point predicts each person's next measurement as their last observed value. As a second naive reference, it uses the cohort-mean trajectory. Any model that cannot beat carrying the last value forward over one-year gaps is not learning dynamics.
The strongest classical competitor reproduces the Taiwan cohort's approach: residential annual PM$_{2.5}$ as a time-varying covariate in a linear mixed model for trajectories, and in Cox regression for mortality. We will first check that our implementation recovers the published HR of 1.15 (95% CI 1.10–1.20). If it does not, the rest of the comparison is not credible. This baseline is the strongest because it is what the field actually uses, and it has few parameters to overfit to local artifacts.
The strongest simple competitor adds a distributed-lag exposure term and a visit-count covariate to the same mixed model. This is the model a skeptical epidemiologist would build. It gets the same exposure data, the same tuning budget, and the same transport policies as our model.
The strongest continuous-time competitor is a Neural CDE with the observation-count channel from Kidger et al., trained without random effects. As a published reference point for implementation, it reported AUC 0.880 on PhysioNet sepsis. We will reproduce that number before using it on cohort data. We also run ME-NODE with exposure added as an input. For trajectory prediction on any pollution-health cohort, we found no reported result on Lacuna, so every number in these comparisons will come from our own runs.
All models get the same hyperparameter-search budget, measured in GPU-hours, on derivation data only. Target data are never used for tuning.
Two ablations isolate the key ingredient. The first removes the separate visit module and falls back to a single observation-count channel. The second fixes the exposure module at the derivation city's version instead of re-estimating it. Beating the baselines means two things. Under modular transport, the external error must be significantly lower than the full-lock result of the same backbone. The pollution coefficient estimated independently in each cohort must also differ less than it does for the classical mixed model. Winning only on internal splits counts as failure.
Why not just…?
Why not just fit a distributed-lag mixed model with a visit-count covariate? This is the real threat. With annual screening visits, there may be too little information about fine-grained dynamics for a neural vector field to add anything. That is why the modular design is built to accept this model as a backbone. If the lagged mixed model inside the modular design matches the Neural CDE version, the conclusion is that the decomposition helps transport and the network does not, and we will report exactly that. What we will test is whether a single-covariate visit adjustment corrects the bias from sicker people being seen more. We test it in the semi-synthetic stage, where the true effect is known.
Why not just impute onto a regular grid and run a GRU or TimesNet? On CharacterTrajectories with 70% of values missing, Neural CDEs reached 98.6% accuracy against 91.9% for GRU-D. Under heavy noise on MIMIC, EDICT kept AUROC near 0.80 while the discrete baselines fell to 0.50. Those results come from short ICU or handwriting sequences, not multi-year cohorts. We therefore include a grid-imputed GRU and a TimesNet baseline rather than assume they lose.
Why not just fine-tune the whole model on the target cohort? Fine-tuning needs labeled health outcomes from the target, and it replaces the dose-response estimate with a local one. That defeats the purpose of learning whether the dose-response transports. We plot error against the number of labeled target people for full refit. The modular policy should match full refit at small target sizes. If full refit wins at every size, the claim fails.
Why not just use leave-one-region-out cross-validation within Taiwan? We run it, and we measure whether it predicts the external gap. The OHDSI result shows that shifts in how data are recorded, like the drop to AUC 0.50 on the German registry, are invisible inside a single data system. If held-out regions within Taiwan reproduce the external drop, that alone would be a useful and cheaper finding.
Proposed Experiments
Stage one: semi-synthetic, where the true effect is known. We use the Xi'an fleet maps from the taxi-fleet study as a realistic street-level exposure field. We simulate people with home-work routes and walking or cycling modes, applying the 4–9× ventilation multipliers from the e-bike study. Each simulated person gets a latent health state with a known nonlinear dose-response and person-specific effects. Visits are drawn from a rate that rises with the latent state.
We create a synthetic "second cohort" by changing two things independently: sensing (taxi coverage versus station-only sub-sampling) and visit policy (fixed annual versus illness-driven). Because we control each shift separately, we can attribute error to one module at a time. The decisive check has two parts. First, modular transport should recover the true dose-response coefficient within its confidence interval under both kinds of shift, while full lock should be biased under visit-policy shift. Second, removing the visit module should reproduce that bias.
The limitation is plain. The two-week Xi'an window is recycled to cover multiple years, so this stage tests identification, not biology.
Stage two: real cohorts. We derive the model on the MJ Health cohort, using repeated screening measures, residential exposure, and questionnaire activity converted to ventilation. Validation is in two layers. The first is a held-out-region split within Taiwan. The second is an external biobank cohort from the adiposity study, either UK Biobank or the Estonian Biobank. The external layer depends on linking residential PM$_{2.5}$ estimates to those cohorts, which the file does not establish. If the linkage is not possible, stage two reduces to the within-Taiwan layer, and we state that the external claim remains untested.
Metrics and decisive check. We measure error on each person's next measurement, calibration of prediction intervals, and C-index for mortality. We compare the pollution coefficient estimated in each cohort. As a secondary estimand, we check whether modeling dose with ventilation changes the small exercise-by-pollution interaction reported in Taiwan (HR 1.03), since coarse exposure may be hiding it. The decisive result is whether modular transport closes most of the gap between internal and external performance at zero or few target labels, and whether the gain survives when the backbone is the simple lagged mixed model.
References
- Development and validation of a prognostic model predicting symptomatic hemorrhagic transformation in acute ischemic stroke at scale in the OHDSI network
- Effects of air pollution and habitual exercise on the risk of death: a longitudinal cohort study
- Evaluating Students’ Dose of Ambient PM2.5 While Active Home-School Commuting with Spatiotemporally Dense Observations from Mobile Monitoring Fleets
- A Context-Aware E-Bike System to Reduce Pollution Inhalation While Cycling
- Mixed Effects Neural ODE: A Variational Approximation for Analyzing the Dynamics of Panel Data
- Neural Controlled Differential Equations for Irregular Time Series
- Random Controlled Differential Equations
- Characterising the genetic architecture of changes in adiposity during adulthood using electronic health records
- Continuous Time Evidential Distributions for Irregular Time Series