Individualized Prediction of Migraine Attacks Using a Mobile Phone App and Fitbit

Eliminating migraine attacks before they start is of an enormous importance to migraine
sufferers. But figuring out the onset of an attack before it actually starts remains a major
challenge for the medical community.

The widespread use of mobile smartphones, the availability of wearable devices that measure
health information, and advances in multivariate pattern analysis via machine learning
algorithms allow for development of individual predictive models that can determine the
likelihood of an individual patient developing a migraine on a given day. Such models are
based upon objectively measured biometric parameters (e.g. activity, sleep), objectively
measured environmental conditions (e.g. weather parameters), exposures to possible migraine
triggers, and patient reported symptoms. Using machine-learning algorithms to explore this
large dataset that is collected for each patient, the optimal combination of factors that
most accurately predict the likelihood of a migraine attack is determined.

Prediction of individual migraine attacks would have substantial positive impacts for
patients with migraine. Accurate prediction of a migraine attack would give the migraineur a
greater sense of control over their condition, a sense of control that is often lacking in
patients with migraine. Most importantly, if individual migraine attacks could be predicted
with high accuracy, treatment of that inevitable migraine attack before development of
symptoms could prevent the attack altogether.

Eligible subjects will enter a baseline phase during which subjects will wear a Fitbit device
and record data into the daily headache diary using the mobile phone app. This phase will be
of variable duration for each subject to a maximum of 75 days. It is during the baseline
phase that the individualized predictive model for a migraine attack is developed and
optimized.

During the second phase (75 days), the accuracy of the predictive model will be tested. The
probability of developing a migraine will be calculated and the accuracy of the prediction
will be tested against the patient reported incidence of migraine attacks within the mobile
phone app. Subjects will be blinded to the app's migraine attack predictions to avoid
expectancy bias.

Migraine prediction suffers from 'the curse of dimensionality' (machine learning parlance).
Too many factors affect outcomes, but the outcomes (positive migraine attacks) are few and
far in between. To develop an accurate machine learning model using traditional approaches
requires a long and impractical time duration. Migraine Alert has effectively addressed these
using proprietary algorithms and techniques that generate individual models using fewer
migraines. Covariate analysis is performed for each individual using features derived from
the raw data. Individual models may differ from one other in the specific feature they use
and/or the importance attached to them in the model. Proprietary techniques are used to
create these individual models and to monitor their pre-validation and post-validation
accuracy and recall. Concept drift as evidenced by any degradation in accuracy or recall is
monitored in the prediction phase and model is retrained as necessary.

Inclusion Criteria:

- Subjects fulfilling ICHD-3beta criteria for migraine with average of 5 - 10 migraine
attacks per month and up to 12 headache days per month

- Males of females 18 years of age or older

- Subject report of weather being one of the triggers

- Subject has an iPhone

- Subject is willing to wear a Fitbit device for the duration of the study

- Subject has an active Facebook account or is willing to create one

Exclusion Criteria:

- Children younger than 18 years of age

- Subjects with headaches other than migraine or probable migraine

- Inability to provide informed consent

- Not willing to maintain a daily diary

- Current participation in another clinical trial