- A forecast predicts encounters, such as emergency visits, clinic visits and admissions for flu-like illness, by region and by week, for a horizon you choose.
- H2O Driverless AI builds time series models from surveillance counts, lab results, calendar effects and other signals you already collect.
- The same approach tracks vaccination progress by region and helps reviewers sort free-text adverse event reports faster.
- Forecasts inform planning; your epidemiologists and program leaders make every decision, and H2O makes no clinical efficacy claim.
Every season arrives on its own schedule
Respiratory seasons do not start, peak or end on the same week each year. When your plan rests on last season's curve, you find out late that clinics are full, staff are short, or vaccine stock sits in the wrong county.
Your surveillance data shows what happened last week. Planning needs a view of the next few weeks, region by region, that updates as new counts arrive.
How public health forecasting works
Disease forecasting is a predictive AI application. It uses time series models trained on your own history to predict future counts from past counts, the calendar, and the signals that move ahead of them, built in H2O Driverless AI inside your own boundary.
Step 1: connect the history
Pull weekly encounter counts, lab positivity and syndromic data by region from the systems you already run.
Step 2: add the signals
Include calendar effects, school terms, weather and any early indicators your epidemiologists trust.
Step 3: build and compare models
H2O Driverless AI creates lag, rolling and calendar features through automatic feature engineering and ranks candidate models on held-out weeks. Your team can add its own epidemiological model as a custom recipe.
Step 4: explain the forecast
ML interpretability shows which signals moved each forecast, so your analysts can check it against what they see in the field.
Step 5: refresh and monitor
Retrain as new weeks arrive, and watch for model drift when a new strain or a change in testing shifts the pattern.
Vaccination programs and adverse event review
The same forecasting approach predicts vaccination uptake by region and age group, so a program can see where coverage will fall short and move outreach and stock early.
Natural language processing can also sort free-text adverse event reports submitted after vaccination by the terms they contain, so safety reviewers read the most relevant reports first. Reviewers make every assessment; the model only orders the queue.
Most forecasts run on counts that are aggregated before modeling. When a model does use patient-level records, protected health information never has to leave your boundary.