Can AI Predict Health Impacts from Wildfire Smoke and Flooding?

AI for predicting health impacts of wildfire smoke and flooding is strongest today where the endpoint is still environmental: where smoke will travel, how deep water may get, which counties may see a surge in exposure. It is much weaker where the endpoint becomes clinical: which child with asthma will need escalation, which post-flood patient will develop an infection, which service line should change treatment before symptoms appear. As of Q3 2026, there is no FDA-cleared or CE-marked AI device for predicting individual health impacts from wildfire smoke or flooding.

That distinction matters because the environmental layer is no longer speculative. Several systems now forecast smoke, emissions, floods, and elevation-related flood risk at a level that can support public health preparation. The clinical layer, by contrast, still has a thin evidence base, limited external validation, and no settled regulatory pathway for patient-level decision support.

The Forecasting Layer Is Getting Hard To Ignore

Wildfire smoke forecasting has moved well past static maps and retrospective air-quality summaries. In January 2026, CIRES and NOAA described an AI system designed to predict wildfire emissions 35 to 45 days ahead by combining seven global fire-emission inventories with meteorological and vegetation data.[1] That is not a bedside model, but it is meaningful for preparedness: a health department, an asthma clinic, or a pediatric respiratory service can do more with a credible weeks-ahead emissions signal than with a plume map that arrives after the waiting room is full.

Shorter-range smoke tools are closer to operational response. The University of Utah’s Trace AQ tool, announced in August 2025, uses physics-based modeling with machine-learning augmentation to forecast wildfire smoke four days in advance.[2] Microsoft’s Aurora model, as described by Harvard Medicine, calculates global air-pollution patterns five days ahead.[3] A hybrid machine-learning and chemical-transport-model approach reported air-quality prediction accuracy above 90%, outperforming either component alone by 66% and 12%, respectively.[4]

The useful development here is not that machine learning has replaced atmospheric science. It is that hybrid systems can absorb physical constraints, satellite observations, weather inputs, vegetation signals, and historical emissions patterns into forecasts that arrive early enough for action. A hospital does not need a smoke model to diagnose asthma. It may need a smoke model to decide whether to expand clinic slots, move vulnerable patients indoors, coordinate medication refills, or warn school-based health teams before smoke peaks.

Flood AI Is Following A Similar Operational Path

Flood forecasting shows the same pattern: increasingly capable environmental intelligence, not yet clinical prediction. Google says its Flood Hub provides machine-learning-powered riverine flood forecasts up to seven days in advance across 80 countries, covering 460 million people. The same Google update also describes a wildfire boundary tracker that updates every 15 minutes using AI and satellite imagery.[5]

At a more local scale, Georgia Tech described a physics-informed AI system in June 2026 that forecasts building-level flood depths three to five days before hurricane landfall with greater than 90% accuracy, with the explicit goal of helping communities stage resources before a storm.[6] Fathom’s 2026 discussion of FathomDEM+ points to another bottleneck: elevation data. Its hybrid AI approach uses vision transformers and convolutional neural networks trained on LiDAR data spanning more than 10 million square kilometers to correct global elevation data for flood modeling.[7]

Those systems answer operational questions that matter before clinical questions are even reachable. Which buildings may flood? Which roads may be impassable? Which neighborhoods may lose access to dialysis, pharmacies, urgent care, or clean water? Which shelters need medical staffing? These are not minor proxies. They are the conditions under which health risk emerges.

The Missing Step Is Not More Maps

The hard part is the chain from hazard forecast to health outcome. A smoke plume forecast estimates where particulate exposure may occur. A flood-depth model estimates where water may accumulate. Neither one, by itself, knows who lives there, who has uncontrolled asthma, who lacks transportation, who depends on powered medical equipment, who will be exposed indoors after mold growth, or who will present to an emergency department rather than call a primary care office.

For clinical use, the model would have to cross several thresholds. It would need to infer exposure at a clinically meaningful level, connect that exposure to patient or population vulnerability, predict an outcome that changes action, perform well outside the development setting, and fit into a workflow where someone is responsible for reviewing and acting on the alert. A forecast that smoke will reach a city in four days can justify preparedness. A recommendation that a specific patient should receive prophylactic medication requires a different evidentiary standard.

This is where the category error often enters. Environmental forecasting tools can be impressive without being clinical decision support. They can support command centers, public health dashboards, staffing decisions, and outreach lists. But unless they have been validated against clinical endpoints and deployed through a regulated pathway where appropriate, they should not borrow the authority of medical AI.

Health Outcome Evidence Is Still Thin

The small evidence base is the first reality check. A 2024 PLOS Climate review, cited by Harvard Medicine, identified only seven English-language studies that used machine learning to directly predict health outcomes from climate-driven events.[3] That number is striking because it sits beside a much larger and more mature body of environmental forecasting work. The imbalance is the field’s central problem.

One of the more relevant examples comes from Boston Children’s Hospital. Harvard Medicine describes work by John Brownstein and colleagues in which a machine-learning model predicted regional pediatric respiratory capacity needs “to the day” during the 2023 Canadian wildfire season.[3] That is a genuinely useful clinical-adjacent endpoint. It tells a pediatric respiratory service something it can act on: capacity, staffing, bed pressure, and readiness for smoke-associated demand.

It is also not the same thing as predicting that a particular child will deteriorate on a particular day. Capacity forecasting can succeed with aggregate signals: smoke, seasonality, regional utilization, prior visit patterns, and service demand. Individual risk prediction has to survive messier inputs, including diagnosis quality, medication adherence, housing conditions, indoor filtration, access to care, and the timing of actual exposure. The operational value of the Boston Children’s example is high precisely because it stays near a decision hospitals can make now.

Flood-Related Disease Prediction Shows Both Promise And Fragility

Flood-related health prediction is similarly early. A 2025 retrospective post-flood infectious-disease study with 939 patients found infectious-disease prevalence increasing from 39.5% before flooding to 47.3% after flooding, with an odds ratio of 1.38. In the same study, a Random Forest model achieved an AUC of 0.76 for infectious-disease prediction.[8]

Those numbers are useful, but the feature behavior limits the clinical interpretation. Age and visit date dominated model performance, while comorbidities and gender contributed minimally; the study also found younger adults at higher risk than older adults.[8] That may reflect true local exposure patterns, care-seeking behavior, occupation, cleanup activities, or dataset structure. A single retrospective study cannot settle which explanation is right.

For an infection-control team, this kind of model may still be worth watching. It can suggest when post-flood surveillance should intensify and which broad groups may need attention. For a clinician deciding whether to treat an individual patient differently, the same evidence is not enough. AUC 0.76 in one retrospective setting is not a license to automate prophylaxis, triage, or discharge decisions.

Disease Forecasting Is Advancing, But It Does Not Collapse The Gap

Broader infectious-disease forecasting research is more encouraging. A 2025 Springer Nature review reported that AI approaches, including deep learning and hybrid models, significantly outperform traditional compartmental models for climate-driven disease forecasting, while also noting data scarcity in low-resource regions.[9] That supports the general direction of travel: models that combine climate, mobility, epidemiology, and other signals can improve outbreak forecasting.

Johns Hopkins described PandemicLLM in June 2025 as an LLM-based reasoning system using demographic, epidemiological, policy, and genomic data to predict infectious-disease trends one to three weeks ahead, outperforming CDC CovidHub benchmarks.[10] That is relevant because it shows how reasoning systems may combine heterogeneous signals, not because it proves wildfire-smoke or flood-related bedside prediction is ready.

The same caution applies to platform ambitions. A model that predicts disease trends across a region can inform staffing, testing supplies, public messaging, and surveillance intensity. It does not automatically identify which exposed patient should receive an intervention. Population forecasting and patient-level clinical prediction share data sources, but they do not share the same validation burden.

What A Hospital Can Reasonably Do With These Models Now

The near-term use case is preparedness, not autonomous clinical decision-making. A hospital or public health agency can use smoke and flood forecasts to adjust staffing assumptions, prepare respiratory clinics, review high-risk registries, stage mobile units, coordinate pharmacy access, and anticipate disruptions in transportation or utilities. Those decisions tolerate uncertainty better than individual treatment recommendations because they are operational and reversible.

• Use environmental forecasts as situational awareness, not as patient-specific diagnosis or treatment logic.
• Separate exposure alerts from clinical risk scores in governance documents and dashboard labels.
• Validate local associations between smoke, flood conditions, visits, admissions, and capacity before changing workflows.
• Require external validation before using a model outside the geography, population, or event type in which it was developed.
• Treat patient-level recommendations as clinical decision support and assess whether a regulatory pathway applies.

The distinction can be uncomfortable in procurement discussions because the same dashboard may display hazard forecasts, hospital utilization trends, and individual patient lists. But the evidence behind those layers is not equivalent. A flood-depth forecast can be operationally mature while the downstream infection-risk score remains exploratory. A smoke forecast can be reliable enough to trigger outreach while still being insufficient to justify a medication change for one patient.

The Practical Answer In Q3 2026

Current AI systems can increasingly predict wildfire smoke and flood hazards, and some can support population-level health-system preparedness. The strongest examples sit in environmental intelligence: emissions lead time, smoke movement, air-quality forecasting, flood extent, building-level flood depth, and elevation correction. Those outputs can help hospitals and public health agencies act earlier.

Current AI systems cannot yet be treated as validated individual-level clinical risk prediction for health impacts from wildfire smoke or flooding. The direct health-outcome literature is small, the best examples are mostly capacity or disease-surveillance models, and no cleared or CE-marked device exists for this purpose as of Q3 2026. The responsible deployment boundary is therefore narrow: use these tools to prepare services and reduce exposure, but do not present them as regulated bedside prediction unless the clinical endpoint, validation evidence, workflow, and regulatory status actually support that claim.

References

1. Artificial intelligence takes on wildfire emissions, CIRES/NOAA, January 2026.
2. U Scientists Develop AI-powered Tool to Forecast Wildfire Smoke, University of Utah, August 2025.
3. Machine Learning Can Predict the Weather — and Human Health, Harvard Medicine.
4. Hybrid machine learning and chemical transport modeling for air quality prediction, Popular Science / The Conversation.
5. How Google AI helps combat wildfires, floods, and extreme heat, Google Blog.
6. How AI-Powered Flood Forecasts Could Transform Hurricane Resilience, Georgia Tech, June 2026.
7. AI And Machine Learning In Flood Modeling, Fathom, 2026.
8. PMC12592683, PubMed Central, 2025.
9. AI applications in climate-driven disease forecasting, Discover Public Health / Springer Nature, 2025.
10. PandemicLLM, Johns Hopkins Hub, June 2025.