Know your migraine risk before it starts.

The baseline model is built on two sources: published clinical research on environmental migraine triggers — in particular barometric-pressure dynamics, temperature swings, humidity, and air quality — and aggregated, anonymized usage statistics from the Relief user base: population-level patterns of when attacks are logged relative to environmental conditions.

No individual user’s identifiable health records are part of this dataset; personal data never leaves your device, so it can’t be. Your own logs are used only on your device, to personalize the model for you.

From day one you get the population-level baseline forecast: the app evaluates the environmental conditions at your location — above all the rate and magnitude of barometric-pressure change, plus temperature shifts, humidity, and air quality — against patterns known to affect migraine sufferers in general. It’s a “weather-sensitivity” forecast, not yet a personal one.

As you log attacks and triggers, a personalization layer is trained on-device and gradually re-weights the baseline toward your individual sensitivities.

The first personal adjustments appear after roughly ten logged attacks or about a month of regular use, whichever comes first. Confidence grows with data: after around thirty attacks over a few months, the forecast is predominantly driven by your own patterns rather than the population baseline.

There’s no hard cutoff — personalization is continuous, and each new entry refines the weights.

At the population level, the strongest predictors are, in order: rapid barometric-pressure changes (drops especially), large temperature swings within 24–48 hours, humidity extremes, and poor air quality (PM2.5 and ozone).

For a personalized model there’s no universal answer — weights are individual. Some users turn out to be pressure-sensitive, others cycle-driven or sleep-driven, and the app’s statistics screens show you which factors rank highest for you.

Not only weather. The model can incorporate every signal available on your device: weather and air quality (always, if location is granted); menstrual cycle phase, heart rate, and heart rate variability from Apple Health (if you grant those permissions); and everything you log manually — sleep quality and duration, stress, diet, physical activity, screen time, medications, and prior symptoms.

The more channels you enable, the more complete the picture — and all of this fusion happens on-device.

Through temporal-window analysis, similar to a case-crossover design: the app compares how often a factor — poor sleep, a specific food, an aura symptom — occurred in the 24–48 hours before your attacks versus comparable attack-free periods in your own history.

A factor is flagged as a likely trigger or precursor only when it appears before attacks substantially more often than its baseline frequency in your life. In other words, the app measures an effect size against your own control periods, not just co-occurrence.

Several safeguards work together: a minimum-occurrence threshold (a pattern must repeat across multiple independent episodes before it is surfaced — one coincidence is never enough); comparison against baseline frequency (chocolate “before an attack” means nothing if you eat it daily); statistical-significance testing of the association strength; and confidence grading — insights are shown with an indication of how strong the evidence is, and weak associations are either hidden or explicitly labeled as tentative.

With small histories the model deliberately stays conservative and leans on the population baseline.

Continuously, on two triggers: whenever fresh weather and air-quality forecast data arrives from our backend — typically every few hours — and immediately whenever you log new data or new Apple Health samples become available. The risk on the home screen therefore reflects both the latest environmental forecast and your latest entries.

Honestly: it’s a probabilistic estimate, not a prediction of fact, and accuracy is inherently individual — highest for people whose attacks are strongly environment- or cycle-linked, lower for those with irregular, trigger-independent migraines. Accuracy improves measurably as your on-device history grows.

The forecast has not undergone clinical validation and is not a medical device — treat it as an informational aid, not a diagnosis.

They’re an unavoidable part of any probabilistic system, in two forms: false alarms (elevated risk, no attack) and misses (an attack on a “calm” day). The model is deliberately tuned toward sensitivity — we consider a false alarm less costly than a missed warning — so expect occasional high-risk days that pass uneventfully, especially in your first weeks. Misses typically reflect triggers the model cannot see, like an untracked factor or an unusual stressor.

Both rates decline as your history grows — and logging attacks even when no warning was shown is the single best way to reduce them, because those entries correct the model.

Core meteorological data — pressure, temperature, humidity, wind — is model-based and covers effectively any coordinates, so the weather forecast works in every country. Air-quality and pollutant data depends on ground-station and satellite coverage: it’s densest in regions with extensive monitoring networks, such as North America, Europe, and parts of Asia-Pacific, and may be estimated or occasionally unavailable where networks are sparse.

Where a specific layer is unavailable for your location, the forecast simply runs without it — the app remains fully functional everywhere.