Why Outdoor Wildfire Smoke Monitors Miss the Real Health Impact

The central measurement problem in wildfire smoke research is almost embarrassingly practical: the monitor is outside, while the person is usually not. North Americans spend roughly 90% of their time indoors, yet wildfire smoke epidemiology has commonly assigned exposure from outdoor PM2.5 estimates, as if the air at a fixed monitor were the air entering a resident’s lungs, a care-facility room, or a child’s bedroom during a smoke episode.[1]

That shortcut matters for wildfire smoke effects on public health and event cancellations because decisions often begin with outdoor air-quality numbers. Outdoor PM2.5 is useful for surveillance and public communication, but the British Columbia study shows why it is too blunt when the question is health-effect estimation. A 10 µg/m³ increase in machine-learning-modeled indoor PM2.5 was associated with a 10.1% increase in salbutamol dispensations, an asthma rescue medication. Four outdoor PM2.5 metrics showed only 3.6% to 6.1% increases, with no overlap in 95% confidence intervals.[1]

That is not a small technical adjustment. It changes the apparent size of acute respiratory harm by about 1.5 to 3 times compared with outdoor metrics. For this acute respiratory indicator, outdoor-only exposure estimates appear to bias the estimated effect toward the null.[1]

Why Outdoor PM2.5 Pulls the Estimate Down

Exposure misclassification is not a statistical nuisance here; it is the main event. If a study assigns everyone in an area the same outdoor smoke level, it treats a resident in a tightly sealed, air-conditioned building like a resident in an older, leakier building with open windows or limited filtration. It treats a care-facility room like a sidewalk. When those assigned exposures are wrong in ways that blur high and low indoor exposure, the estimated health effect tends to be diluted.

During wildfire smoke episodes, indoor PM2.5 is partly driven by outdoor smoke, but it is not simply outdoor smoke divided by a constant. Particles enter through ventilation systems, gaps, doors, windows, and human behavior. Filtration, air conditioning, building age, maintenance, and room-level conditions all intervene before smoke becomes inhaled exposure. The BC study is valuable because it attempts to model that intervening layer rather than pretending it is invisible.[1]

This is especially relevant for populations whose risk is easy to flatten into area averages. Residents in care facilities may spend nearly all of a smoke episode indoors, and their exposure depends heavily on the building that surrounds them. People with asthma may not appear in a hospital dataset but may still require more rescue medication. Local officials deciding whether to cancel outdoor events or modify services are often looking at outdoor air-quality dashboards; health researchers, however, need to know whether those dashboards are measuring the biologically relevant exposure well enough.

What the British Columbia Model Actually Did

The study trained a weighted machine-learning ensemble on colocated indoor and outdoor low-cost PM2.5 sensors at 44 care facilities in British Columbia during the 2022 and 2023 wildfire seasons. The ensemble used random forest, quantile regression forest, and xgBoost models to estimate indoor PM2.5 at population scale.[1]

The important point is not that these algorithms are fashionable. It is that they were assigned a job epidemiology badly needs: infer indoor exposure where direct indoor monitoring is not available for everyone. On test data, the weighted ensemble achieved an RMSE of 3.29 µg/m³ and an R² of 0.71. In the extreme 2023 wildfire-season validation comparison, the indoor model had an RMSE of 6.65, compared with 9.64 for the outdoor-only model.[1]

The 2023 validation result matters because wildfire smoke models can look adequate under ordinary conditions and then fail when they are most needed. An extreme season tests whether the relationship between outdoor smoke and indoor air remains predictable when concentrations and infiltration patterns are stressed. The BC model did not make indoor exposure perfectly observable, but it reduced error compared with relying on outdoor PM2.5 alone.[1]

Indoor Exposure Follows Buildings and Social Conditions

One of the more consequential findings is that indoor smoke exposure had a social and infrastructural pattern. Adding Canadian Index of Multiple Deprivation variables—economic dependency, situational vulnerability, ethno-cultural composition, and residential instability—meaningfully improved the model’s indoor PM2.5 predictions. In other words, the same wildfire event did not translate into the same indoor exposure across communities.[1]

That should make public-health readers cautious about averages. Deprivation variables are not decorative covariates; they stand in for conditions that can shape the air people actually breathe, including housing quality, crowding, resources for filtration, and the ability to keep windows closed during heat and smoke. The study does not turn those variables into a complete causal map, but it shows that excluding them worsens exposure prediction.[1]

Regional infrastructure showed the same pattern. Interior BC regions with high air-conditioning prevalence, including Kelowna at 84%, had systematically lower indoor/outdoor PM2.5 ratios than coastal Vancouver, where air-conditioning prevalence was 26%. That contrast is a reminder that indoor exposure is moderated by built environments, not only by plume intensity.[1]

This is where the machine learning earns its keep. A simpler outdoor-only model can tell researchers that a region was smoky. It cannot tell them, with the same specificity, how smoke was filtered through buildings, cooling infrastructure, and deprivation-linked conditions before it became indoor exposure. For health-effect studies, that difference can decide whether the estimated effect looks modest or clinically harder to ignore.

The Health Signal Is Concrete, but Bounded

Salbutamol dispensations are a useful acute respiratory indicator because they reflect a specific response: more people needed rescue medication. The BC study found that modeled indoor PM2.5 had a stronger association with this outcome than any of the outdoor PM2.5 metrics. It also reports this as the first study to quantify the relationship between modeled population-scale indoor PM2.5 and an acute respiratory health indicator during a North American wildfire season.[1]

The finding should not be made larger than it is. It does not say that every smoke-related health outcome is underestimated by the same factor. It does not replace clinical assessment for individual patients. It does not prove that indoor PM2.5 estimates are equally transferable to every US region, where housing stock, air-conditioning prevalence, building codes, and behavioral adaptations may differ from British Columbia.

It also depends on low-cost PM2.5 sensors rather than reference-grade monitors, so calibration, placement, and maintenance matter. That is not a reason to dismiss the study; population-scale indoor exposure assessment would be nearly impossible if it required reference-grade indoor monitors everywhere. It is a reason to treat sensor quality and deployment design as part of the evidence, not an implementation detail.

Why Better Exposure Estimates Matter Beyond One Dataset

The broader literature gives this measurement problem additional weight. Stanford research reported that wildfire PM2.5 can be about 10 times more toxic per unit mass than ambient fossil-fuel PM2.5, a finding that makes exposure accuracy more than a technical preference.[2] If the pollutant mixture is unusually harmful, then assigning the wrong exposure can meaningfully distort risk estimates.

Harvard Chan School coverage of wildfire smoke research has also emphasized that cardiorespiratory hospitalization risks can persist for 3 months after a fire has ended.[3] That longer time horizon is not the same outcome as salbutamol dispensations, and it should not be folded into the BC estimate. But it does explain why researchers should care about acute exposure estimates: the first measurement error can propagate into downstream analyses of care demand and delayed health effects.

For event cancellations, school activities, athletic practices, and public gatherings, outdoor PM2.5 will remain the practical starting point. It is visible, standardized, and communicable. The BC study does not supply a decision rule for canceling an event, nor does it claim to forecast smoke movement. It does, however, warn against assuming that the outdoor number fully represents exposure for the population affected by those decisions.

That distinction matters for health IT and environmental health teams. AI smoke forecasting tools address where smoke may go; indoor exposure models address what people may actually inhale after smoke reaches a community. Those are neighboring problems, not interchangeable ones. For a broader discussion of models that try to predict health outcomes from smoke, see How Well Does AI Predict Wildfire Smoke Health Effects? The BC paper is best read as a methodological advance in exposure assessment, not as a real-time protection system.

What Changes When the Person Replaces the Monitor

Once indoor exposure is modeled, the health-effect estimate changes. That is the result that should hold attention. The study does not merely show that machine learning can fit sensor data with acceptable accuracy; it shows that a better exposure metric produces a materially larger association with acute respiratory medication use.[1]

Outdoor PM2.5 remains indispensable for public dashboards, alerts, surveillance, and many smoke-related operational decisions. It is still the number most communities can see quickly. But as the sole exposure metric in health-effect estimation, it is too crude. It misses the buildings, infrastructure, and social conditions that determine how much smoke reaches indoor air.

The measured conclusion is narrower than an AI headline and more useful than one: this study does not make machine-learning smoke forecasting claims, and it does not provide individual clinical guidance. It shows that when researchers model the air people likely breathe indoors, the respiratory health burden of wildfire smoke looks larger than outdoor monitors alone suggest.

References

1. Enhancing Wildfire Smoke Exposure Assessment: A Machine Learning Approach to Predict Indoor PM2.5 in British Columbia, Canada — ACS Environmental Science & Technology Air.
2. Assessing wildfire health risks — Stanford News, 2025.
3. Cardiorespiratory effects of wildfire smoke particles can persist for months even after a fire has ended — Harvard T.H. Chan School of Public Health.