How AI models forecast respiratory disease from air quality

The practical question behind AI in air quality and respiratory health forecasting is not whether a model can draw a clean curve after the fact. It is whether a pediatric emergency department, asthma clinic, or hospital operations team can see enough of the next respiratory surge to adjust staffing, appointment capacity, messaging, or surveillance before patients are already waiting.

That is why the São Paulo pediatric emergency department study is a useful place to start. Cabral-Miranda and colleagues analyzed 24,366 pediatric respiratory visits and trained an XGBoost model using clinical, pollution, and climatic factors, including PM10, ozone, temperature, and humidity. The model predicted pediatric respiratory emergency care with 85% accuracy, reported an AUC of 0.81, and forecast daily case volumes 5 to 30 days ahead.[1]

Those details matter. The outcome was not a hidden biological state inside one child’s airway. It was service demand: how many children with respiratory disease were likely to arrive on a given day. The lead time was also operationally meaningful. A five-day signal can affect staffing rosters, respiratory therapist allocation, medication stock checks, and outpatient overflow planning. A 30-day signal is less likely to be precise enough for day-by-day decisions, but it can still support seasonal readiness if it is stable in local validation.

What the stronger studies actually predict

The most credible near-term use case is population-level forecasting: emergency department visits, outpatient asthma volume, or daily respiratory clinic load. These are noisy outcomes, but they are the kind of outcomes hospital teams can act on. They also match the structure of environmental data. Pollution, weather, pollen, and influenza activity affect groups unevenly, but their signals are usually measured at neighborhood, city, regional, or national scale.

The South Korean asthma study widened the question from pediatric emergency demand to asthma care volume. Hwang and colleagues used five years of national data and compared LSTM against generalized linear models, generalized additive models, Random Forest, and Gradient Boosting. The LSTM model performed best for predicting asthma outpatient visits, with R²=0.723, and emergency room visits, with R²=0.650, using 18 environmental factors that included influenza, PM10, NO2, CO, pollen, and temperature.[2]

The important point is not that one algorithm has been crowned for every respiratory forecasting problem. It is that a temporal model can be well matched to a temporal clinical burden. Asthma visits do not rise only because today’s pollutant value crosses a line. They can reflect lagged exposure, compounding weather conditions, circulating influenza, pollen peaks, and care-seeking behavior that moves on its own schedule. The South Korean analysis specifically found that LSTM captured nonlinear, lagged interactions that more traditional regression approaches missed, and that it showed the least overfitting among the nonlinear models tested.[2]

Tianjin adds a useful caution against treating deeper models as the default answer. Yang and colleagues studied 414,887 respiratory outpatient visits and reported that Random Forest models using temperature, ozone, SO2, and PM10 explained more than 80% of daily respiratory visit variation. SO2 and ozone were the most influential pollutant predictors in that setting.[3]

For a hospital administrator, that is not a lesser finding because the method is less fashionable. A Random Forest model that captures most daily variation in a local outpatient service can be more useful than a more complex model that is harder to maintain, harder to recalibrate, or trained on variables the institution cannot reliably obtain.

The environmental signal is not just pollution

Air quality is central, but the better-performing studies do not treat the air as a single pollutant feed. They combine particulate matter, gaseous pollutants, meteorology, and epidemic or allergen signals. In the South Korean asthma work, influenza and pine pollen emerged as especially important non-pollutant exacerbating factors across the modeled environmental feature set.[2]

That fits the clinical pattern. A week with moderate pollution may not look dangerous in isolation. Add a temperature shift, high pollen, humidity changes, and influenza activity, and the clinic schedule starts to look different. The value of machine learning here is less mystical than advertised: it can keep track of interacting, lagged signals that clinicians recognize but cannot manually reweight every morning across a city.

The public-health need is not abstract. IQAir reported that only 14% of global cities met WHO annual PM2.5 guidelines in 2025.[4] That does not prove a forecasting model will improve outcomes, but it explains why respiratory forecasting from environmental exposure has become more than a research curiosity. Many health systems are planning under chronic exposure conditions, not waiting for rare pollution disasters.

Population forecasting is not individual prediction

The distinction is easy to blur and clinically important. Forecasting that a pediatric ED may see more respiratory visits next week is not the same as telling a specific child with asthma that they will deteriorate next Tuesday. The first claim is supported by aggregate environmental and visit-volume studies. The second requires patient-level longitudinal data, individual exposure estimates, medication adherence, housing conditions, comorbidities, prior exacerbation history, and evidence that the prediction changes care without causing harm.

At the population level, errors are absorbed differently. If a forecast overestimates respiratory demand by a modest margin, a department may carry extra staffing or open capacity that is not fully used. If it underestimates demand, the waiting room fills, triage slows, and families wait longer. Those are real consequences, but they are operational consequences.

At the individual level, the stakes and evidence requirements change. A false high-risk alert may lead to anxiety, unnecessary visits, medication changes, or alert fatigue. A false low-risk signal may reassure a patient who needs earlier care. Even when the same environmental inputs are involved, the validation question is different: not “did the model anticipate aggregate visits?” but “did patient-specific predictions improve decisions and outcomes compared with usual care?”

Personalized frameworks such as AI-Respire are therefore better read as boundary markers than as proof of deployment readiness. They show where the field may want to go, toward individualized respiratory risk estimation, but the evidence base described for that approach is far smaller than the city- or service-level forecasting studies. The current center of gravity remains demand forecasting, not bedside diagnosis.

Why geography limits the claims

The Brazil, South Korea, and China studies are valuable because they are independent examples, not because they are interchangeable. São Paulo’s pollution mix, pediatric care pathways, climate, and ED access patterns are not the same as Seoul’s national asthma reporting environment or Tianjin’s outpatient respiratory service patterns. A model trained where ozone and SO2 behave one way may not transfer cleanly to a city where wildfire smoke, indoor heating, Saharan dust, or traffic-related NO2 dominates.

Outcome definitions also differ. “Pediatric respiratory ED visits,” “asthma outpatient visits,” “asthma ER visits,” and “respiratory outpatient visits” sound adjacent, but they are not the same endpoint. They capture different ages, diseases, access points, coding practices, and thresholds for seeking care. Cross-study comparisons between LSTM, XGBoost, and Random Forest are therefore weaker than they first appear.

That does not make the studies fragile. It makes them local. The defensible conclusion is that several AI and machine-learning approaches have shown clinically meaningful retrospective performance when trained and tested in their own settings. The less defensible conclusion would be that one can take a published model from São Paulo, South Korea, or Tianjin and expect the same performance in Chicago, Mumbai, Berlin, or Johannesburg without local recalibration and prospective testing.

What LSTM, XGBoost, and Random Forest contribute

A short methods distinction is enough for the clinical reader. LSTM models are designed for sequential data, which makes them plausible candidates when yesterday’s influenza activity, last week’s pollen, and several days of pollutant exposure may all shape tomorrow’s asthma visits. XGBoost is a gradient-boosted tree method that often performs well on structured tabular data with nonlinear relationships. Random Forest builds many decision trees and averages their output, which can make it robust for messy environmental and clinical-volume data.

None of those descriptions proves clinical utility. They explain why the tools fit the problem. Respiratory demand is lagged, nonlinear, seasonal, and locally patterned. Models that can handle interactions among pollutants, temperature, humidity, pollen, and infection activity deserve evaluation. The published results suggest they can produce useful retrospective forecasts, but the model class is only one part of the implementation problem.

The evidence-to-deployment gap

All three main studies are retrospective. That is the right starting point, but it is not the finish line. Retrospective accuracy tells us a model found signal in historical data. It does not tell us whether the forecast changed staffing, reduced wait times, prevented admissions, improved medication timing, reduced exacerbations, or avoided unintended consequences.

A prospective deployment study would need to answer more prosaic questions. Who receives the forecast? How early? At what threshold does the service act? Is the action a staffing change, patient outreach, inhaler refill campaign, school health alert, or clinic capacity adjustment? What happens when the model is wrong for a week? Who monitors drift when pollutant sources, coding patterns, viral surveillance, or patient behavior changes?

This is the same bottleneck seen across medical AI: good retrospective discrimination or fit often arrives long before prospective evidence of improved care. That gap is discussed more broadly in ClinicalMind’s article on why most medical AI studies never reach clinical trials. For respiratory forecasting, the gap is especially visible because the model’s most natural user may be an operations team rather than a single prescribing clinician.

That operational setting may actually be the more plausible first deployment path. A hospital can test whether forecasts improve staffing efficiency, throughput, or surge readiness before asking the model to guide individual treatment. But even then, prospective monitoring is necessary. A forecast that looks excellent across a historical season may fail during an unusual viral year, a wildfire period, a surveillance disruption, or a change in triage policy.

Regulation has not caught up to this use case

As of Q3 2026, no FDA-cleared or CE-marked tool exists specifically for AI-based air quality and respiratory health forecasting. That absence matters less if the model is used as internal situational awareness and more if it starts driving patient-facing alerts, triage decisions, or clinician recommendations.

The regulatory question is not merely whether the software uses AI. It is what the software does in the clinical workflow. A dashboard that forecasts tomorrow’s pediatric respiratory ED volume for staffing may be treated differently from a tool that tells a parent, patient, or clinician that a specific person is at high risk of an asthma exacerbation and should change treatment. The more directly the forecast influences patient-specific clinical decisions, the more likely software as a medical device and clinical decision support classification questions become central.

Those issues overlap with the broader policy questions in ClinicalMind’s analysis of the FDA’s 2026 CDS guidance for AI clinical decision support. For this respiratory forecasting niche, the unresolved point is whether a locally trained, continuously updated environmental-health forecast remains an administrative planning tool, becomes regulated CDS, or crosses into SaMD depending on its output and intended use.

Where the evidence leaves us

AI models can forecast respiratory burden from air quality and environmental data with performance that is meaningful enough to take seriously. The strongest support is for population-level demand: pediatric respiratory emergency visits, asthma outpatient and ER volume, and respiratory outpatient visit counts. The best studies combine pollutants with weather, pollen, and infection signals rather than treating air quality as a single number.

The limits are just as important. The published evidence remains retrospective. Model performance is local to geography, inputs, care-seeking behavior, and outcome definitions. Individual-level forecasting is much less mature than service-volume forecasting. No prospective trial has yet shown that deploying these forecasts improves clinical outcomes, and no dedicated cleared tool exists for this specific use case.

For now, the most defensible reading is practical but restrained: LSTM, XGBoost, and Random Forest models are promising tools for respiratory planning and surveillance when validated locally, but they are not yet proven clinical outcome-improving systems.

References

1. Artificial intelligence platform to predict children's hospital care for respiratory disease using clinical, pollution, and climatic factors — PMC, 2025.
2. Prediction of the number of asthma patients using environmental factors based on deep learning algorithms — PMC, 2023.
3. Prediction of respiratory diseases based on random forest model — Frontiers in Public Health, 2025.
4. 2025 World Air Quality Report — IQAir, 2025.