AAN 2025 | Machine learning models to predict treatment response to migraine preventive medications

So in 2024, our team received the American Headache Society Wolf Award for this particular award and I presented this work as the Wolf Award Lecture at the American Headache Society Scientific Meeting last year. So we utilize a retrospective Mayo Clinic Headache Registry that we have been collecting the detailed clinical presentation and headache characteristic data, as well as the treatment outcome data over the past 20 years. And then we extracted all those data, cleaned all those data, and developed machine learning models based on those clinical features and treatment outcomes that can predict treatment response to seven different types of migraine preventive medications. And that includes commonly used medications like beta blockers, tricyclic antidepressants, topiramate, Botox injection, as well as CGRP monoclonal antibodies, as well as two other oral medications that are commonly used in clinical practice that are verapamil and gabapentin. So throughout this process, we not only develop the models that can accurately predict treatment response to migraine preventive medication, especially for the CGRP monoclonal antibody prediction model, it reached a high AUC, which is typically a parameter indicating model performance. The AUC was 0.83 and an accuracy of 80%. So we not only developed those accurate machine learning models, we also identified several predictors for treatment response or factors that are most important to determine the model, how the model makes the prediction. And those factors include baseline headache frequency, the monthly headache days, patient’s age, patient’s BMI and weight, and other migraine characteristics as well. So for example, whether they have cranial autonomic features and what are patients reported as the triggers for their migraine. And then interestingly, we also identified that those predictors for different medications might be different. So for example, we identified that higher BMI is a positive predictor for Botox injection and topiramate, however, lower BMI is a positive predictor for CGRP monoclonal antibodies. So I think those results indicate that different patients with different clinical characteristics might respond to different medication differently. So that speaks to and emphasizes the need to develop this individualized or personalized treatment plan for migraine because different patients might respond to different medications differently. So I think the important message from this study is that leveraging AI tools, specifically in this study, we use a combination of supervised machine learning with a pre-trained deep learning model and then based on a large clinical registry studies that we collected we developed this machine learning models that can predict treatment response to migraine preventive medications and we identify the predictors for treatment response. So I think overall leveraging AI and large database can help us advance precision migraine treatment that means that we choose the preventive medications based on the individual characteristic of the patient. So the goal is to reduce the trial and error process for a patient to identify the best treatment that is effective and tolerable for each individual.

This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.