A Pragmatic Trial of Machine Prescription for Migraine

Recruiting

• Conditions: Migraine

• Registration Number: NCT06893874
• Lead Sponsor: Norwegian University of Science and Technology

Brief Summary

The purpose of this study is to estimate the ability of ML models to predict the effect of migraine preventives. This will be achieved by first identifying sociodemographic, headache and comorbidity features of migraine patients. Headache days will then be measured for 4 weeks before starting a migraine preventive, and this will be compared to the number of headache days in the first 12 weeks (divided into 28-day periods) after starting treatment. The preventive is regarded as effective if there is a 50% or greater reduction in monthly headache days. After observation of the treatment period, the ML models will use the sociodemographic, headache and comorbidity features, captured before treatment was initiated, to predict treatment effect for all preventives in each participant. These predicted treatment effects will be compared to the actual treatment effects that were observed.

Detailed Description

Rationale As the pathophysiology of migraine is complex, the treatment responses of different migraine preventives are highly heterogeneous. Therefore, migraine patients often try out more than one preventive until an effective treatment is identified. If one can predict individualized treatment effects of these preventives, time to effective treatment may be reduced. We propose that ML may be used to predict individualized treatment effects for migraine preventives (machine prescription).

We are currently developing several ML models for this purpose. The first was developed in 2022 and uses clinical characteristics to estimate individualized treatment effects. The second is under development and uses a combination of genetic and clinical data, while the third will use sociodemographic data. Both of the two latter are trained to predict the effect of commonly used prophylactics in Norway. The result is three different ML models that use different sets of features to predict treatment effects. To compare the performances of these models, we will conduct a pragmatic observational trial to test and compare the models in an independent dataset. Such out-of-sample testing is essential to establish the clinical applicability of the machine prescription models.

Objectives, Endpoints and Estimands The primary objective is to evaluate if machine prescription can predict the treatment effect of commonly used migraine preventives at the individual level. The endpoint for this objective will be the best machine prescription model's ability to predict treatment response (defined as 50% reduction in headache days) measured by ROC-AUC. Secondary objectives include investigating if time-to-treatment-response may be reduced using machine prescription, and if machine prescription more accurately selects the best first-line therapy for migraine at the individual level. Additionally, we will explore developing ensemble models combining the pre-trained models, as well as creating new machine prescription models based on the collected data.

Overall design

This is a prospective observational trial. There will be no control method or blinding. The study population consists of participants with episodic and chronic migraine. Intervention will be standard migraine preventives as prescribed by the physician in care without any interference by the researchers. Each participant will be observed for at least 12 weeks after starting a preventive. Summarized, participation may be presented accordingly:

Screening/inclusion including phone consultation and feature questionnaire (week 0) Baseline (week -4 to 0) Treatment period (week 1 to 12) Follow-up phone call for outcome assessment (week 12) Optional additional treatment period (week 13 to 24) Optional additional follow-up phone call for outcome assessment (week 24)

Brief Summary The purpose of this study is to estimate the ability of ML models to predict the effect of migraine preventives. This will be achieved by first identifying sociodemographic, headache and comorbidity features of migraine patients. Headache days will then be measured for 4 weeks before starting a migraine preventive, and this will be compared to the number of headache days in the first 12 weeks (divided into 28-day periods) after starting treatment. The preventive is regarded as effective if there is a 50% or greater reduction in monthly headache days. After observation of the treatment period, the ML models will use the sociodemographic, headache and comorbidity features, captured before treatment was initiated, to predict treatment effect for all preventives in each participant. These predicted treatment effects will be compared to the actual treatment effects that were observed.

Study details include:

The study duration per participant will be up to 28 weeks. The treatment duration will be decided by the physician in care, but the treatment period in which we will monitor headache days will be 12 weeks for each treatment, with up to two treatment periods for each participant.

The participants will have a phone consultation at inclusion and will have a follow-up phone call at the end of the - or possibly both - treatment period(s) of 12 weeks.

Number of Participants A maximum of 200 participants will be enrolled. Note: Enrolled means participants', or their legally acceptable representatives', agreement to participate in a clinical study following completion of the informed consent process \[and screening\]. Potential participants who are screened for the purpose of determining eligibility for the study, but do not participate in the study, are not considered enrolled, unless otherwise specified by the protocol. A participant will be considered enrolled if the informed consent is not withdrawn prior to participating in any study activity after screening.

Study Arms and Duration There is one study arm. Study duration is described under "Brief Summary".

Data Monitoring There will be no data monitoring committee as it is not applicable for this study.

Study Timeline

• Study Start: 2025-02-01
• Status Verified: 2025-03
• Study First Submitted: 2025-03-18
• Study First Submitted QC: 2025-03-18
• Last Update Submitted: 2025-03-18
• Study First Posted: 2025-03-25
• Last Update Posted: 2025-03-25
• Primary Completion: 2026-12-31
• Study Completion: 2026-12-31

Recruitment & Eligibility

• Status: RECRUITING
• Sex: All
• Target Recruitment: 200

Inclusion Criteria

Aged 18 years or older at trial entry. Episodic or chronic migraine with or without aura as per ICHD-3 criteria. Onset of migraine before age 50 years. History of at least 4 days of migraine in the 4-week baseline period based on retrospective or prospective headache diary assessment or subject recall.

Having recently (within last 4 weeks) started or planning to start a migraine preventive.

Signed and dated informed consent.

Exclusion Criteria

Subject is concurrently using a migraine preventive that is not the drug under investigation.

Subjects who have previously failed all six of the prophylactic treatments we are researching.

Subjects diagnosed with trigeminal autonomic cephalalgias or facial neuralgias. Subjects with secondary headache conditions (except for medication overuse headache according to the ICHD-3).

Subjects taking opioids (>3 days per month) or barbiturates at the time of screening.

Alcohol overuse or illicit drug use. Subjects participating in another clinical investigation.

Study & Design

• Study Type: OBSERVATIONAL
• Study Design: Not specified

Primary Outcome Measures

Name	Time	Method
Prediction of treatment effect	12 weeks	The primary endpoint is to investigate the machine prescription models' ability to predict treatment effect of migraine preventives. To assess this, there will be one binary label of effect versus no effect for each preventive (if a participant has not tried a preventive, the instance will be categorized as missing). The three models will then be applied to the analysis sets, and their label predictions will be compared to the true labels, in a sense using the whole analysis sets as traditional machine learning test sets. The best performing model according to ROC-AUC will be reported.

Secondary Outcome Measures

Name	Time	Method
Time-to-treatment response	12 weeks	The first secondary endpoint is time-to-treatment-response assuming sequential 12-week trials of preventives using machine prescription compared to treatment-as-usual. The predicted effect of each preventive for every participant will be used to generate individualized rankings of preventive effect. Assuming the preventives are given sequentially according to each participant's own preventive ranking, it is possible to estimate how many preventives are expected to be tried before reaching preventive success. This estimation will be compared to treatment-as-usual, which in the study is defined as the preventives participants receive from their physician in care.
Rate of successful first therapy	12 weeks	Ratio of successful first therapy using machine prescription as compared to treatment-as-usual. Each machine prescription model will predict a preventive with the highest probability of effect for every participant. For the participants who received the predicted first choice preventive, the percentage of participants with effect (more than 50% reduction in headache days) will be estimated. For comparison with treatment-as usual, the percentage of participants who had effect of the first preventive given during the study will be estimated. The proportions of successful treatments will be compared with appropriate statistical tests.
External validity	12 weeks	Comparison of external validity and performance, measured by AUC, of the different machine prescription models. AUC will be calculated and presented as a ROC-curve for each model.

Trial Locations

Locations (1)

NTNU Norwegian University of Science and Technology; 🇳🇴 Trondheim, Norway