Machine Learning Assessment of Next-Day Migraine Likelihood

Overview

Development of a machine learning (ML) algorithm for assessment of next-day migraine likelihood, drawing on self-reported migraine-related information, and geographic location, collected via the Nerivio app- a mobile application used for migraine treatment by the Remote Electrical Neuromodulation (REN) wearable device.

The analysis set includes patients with migraine who used Nerivio app for the reporting of migraine attacks and/or associated symptoms and/or migraine-related information. Data collection is through the Nerivio app (Nerivio®). During app registration, patients consent to the collection of de-identified data for research purposes and provide demographic information. Participants can voluntarily report baseline characteristics such as treatment onset time relative to attack onset, headache pain, functional disability, and presence/absence of migraine-associated symptoms, as well as treatment outcomes 2-hours post-treatment. All data is stored on a HIPAA-compliant secure server. The algorithm will be trained on user-level data. The dataset will be structured as a tabular matrix, where the columns represent risk-related features and the rows represent user-day observations. The feature set (X) serves as input variables, while the migraine occurrence label (Y) is the target outcome. The feature set (X) could be divided into four groups, according to the data source: Demographic data - age and sex were self-reported via the app upon registration. The user's country is identified based on IP address. Questionnaire data - features which are the patient's answers to either the daily diary questionnaire or a pre-treatment questionnaire. These include headache severity, functional disability, medication intake, aura, pain duration, and prodromal symptoms. Data on prodromal symptoms is collected via a multiple-choice question with 14 pre-defined answers. Weather data - environmental features that are based on users' geographic location. These include barometric pressure, temperature, heat index, UV index, wind, humidity, and precipitation. Calculated features - features that are calculated based on the aforementioned collected data. These include averages, frequencies, number of consecutive days, etc'. Label (Y) Definition - The target variable (Y) will be defined using daily diary entries and pre-treatment questionnaires. The "next-day migraine" field served as the target outcome. A value of 1 (migraine day) will be assigned if a migraine was reported on the subsequent calendar day; a value of 0 (non-migraine day) will be assigned if no migraine was reported. Missing values were kept as null to capture potential information inherent in their non-random occurrence. A day is classified as a migraine day (Y = 1) if both of the following conditions were met: A) Headache level was reported as mild, moderate, or severe in the diary or pre-treatment questionnaire. B) At least one additional migraine indicator was present: intake of migraine-related medication, or report of photophobia, phonophobia, nausea and/or vomiting, or aura in the diary or pre-treatment questionnaire. The following standard ML outcome measures were used to evaluate model performance: Precision- the percentage of migraine days correctly predicted, out of all predicted days. Accuracy- the percentage of days correctly predicted (migraine and non-migraine), out of all the predicted days. Sensitivity (also termed recall rate)- the percentage of migraine days correctly predicted, out of all the migraine days. Specificity- the percentage of non-migraine days correctly predicted, out of all the non-migraine days. Area Under the Curve (AUC)- the probability that the model ranks a migraine day higher than a non-migraine day, based on predicted risk scores (summarizing the model's ability to distinguish between migraine and non-migraine days across all classification thresholds). F1 score- the balance between correctly predicted migraine days and avoiding wrongly-predicted migraine days, combining sensitivity and precision in one measure.

Conditions

Eligibility

Locations (1)

Outcomes