Methodology
How Pro-Act surfaces migraine risk, and what we do not yet claim.
Our approach
ProAct combines five signals to estimate your migraine risk over the next 48 hours. Each signal is grounded in published migraine research; each is personalized to your history.
Heart rate variability (HRV)
Autonomic nervous system shifts often precede migraine onset. HRV is a non-invasive proxy for these shifts and is captured passively by Apple Watch.
Sleep duration and quality
Both insufficient sleep and oversleeping are well-documented migraine triggers. We use HealthKit sleep stages to detect deviations from a personal baseline.
Resting heart rate
Multi-day elevations or depressions in resting heart rate can correlate with stress load and impending attacks.
Barometric pressure
Sudden changes in atmospheric pressure are a known trigger for many people. We pull local weather data and look for rapid drops or rises against your historical sensitivity.
Self-reported triggers and attacks
Your logged attacks and triggers are the ground truth that personalizes every other signal. Predictions weigh each signal against this history to reflect what matters for you.
How risk estimates work
ProAct estimates your migraine risk through three analysis paths, tried in order. On iPhones that support Apple Intelligence (iOS 26+), the analysis runs entirely on-device using Apple's foundation models -- nothing is sent to our servers for that risk estimate.
When on-device AI is not available, ProAct sends an aggregated summary of your signals -- averages and trends, never raw health samples -- to our backend, where Claude (Anthropic's LLM) evaluates them against your personal attack history. The model returns a structured risk estimate: a risk level for the next 48 hours, the contributing factors that drove that risk, and a confidence band. We do not store this health data on our servers -- the analysis is stateless.
If neither AI path is available, a rule-based engine scores the same signals using weights drawn from published migraine research, so you always get a risk estimate.
Privacy and trust
Health data stays private
Raw health sensor data never leaves your iPhone. On the newest iPhones, predictions run fully on-device; on other devices, cloud predictions use only aggregated metrics, processed in-memory and discarded.
No account required
No email, no password, no personal information -- just an anonymous device ID.
Your logs sync anonymously
Migraine logs sync securely to ProAct's backend under your anonymous device ID -- no account or email, and you can export or delete them anytime.
What we are tracking
We are an early-stage app and have not yet published accuracy results. The metrics below are what we are measuring internally as we grow. We will publish results when we cross the threshold.
Risk-estimate precision
Of all high-risk estimates we issue, what fraction are followed by a logged attack within 48 hours?
Risk-estimate recall
Of all logged attacks, what fraction were preceded by a high-risk estimate in the prior 48 hours?
Per-user trigger inference quality
How accurately does the model identify each user's personal triggers (compared to user-confirmed triggers)?
Calibration
When we say "30% risk" or "70% risk", does that match the empirical attack frequency at that risk level?
We will publish accuracy results once we reach 100+ users with at least 90 days of logged data each.
What we don't claim
- We do not yet have published accuracy results. Specific percentages you may have seen on early marketing surfaces are being removed as we recalibrate to what we can defend.
- ProAct is not a medical device. It does not diagnose, treat, cure, or prevent migraine.
- ProAct does not replace consultation with a neurologist or headache specialist. If your headaches are severe, frequent, or changing in pattern, see a doctor.
- We do not estimate risk for cluster headaches, tension-type headaches, or other primary headache disorders. ProAct is built specifically around migraine signals.
- We do not catch every migraine. Migraines have many triggers, and some attacks happen without detectable warning signs in our current signal set.
Sources
These citations back the approach (signal selection), not specific accuracy outcomes. We will publish our own results when we have meaningful data.
- Koenig J et al. (2016). Heart rate variability in migraine: a systematic review and meta-analysis. Cephalalgia.
- Hoffmann J, Lo H, Neeb L, Martus P, Reuter U (2011). Weather and migraine: a meta-analysis of the literature. Cephalalgia.
- Kelman L, Rains JC (2005). Sleep quality and migraine: a population-based study. Headache: The Journal of Head and Face Pain.
- Andress-Rothrock D, King W, Rothrock J (2010). Triggers of migraine and tension-type headache: experience and knowledge of the patients. Journal of Headache and Pain.