The underlying mechanism of migraine headache is still poorly understood and is likely results due to a combination of genetics and “trigger” events. The patient population is therefore highly diverse and no one treatment fits all migraine sufferers. The extremely debilitating nature of migraine headache, however, demands treatment, which may not be effective.
It follows that 15% of migraine patients surveyed in America, self-reported overuse of migraine medication. Staff at headache clinics report that 50-70% of their clients overuse medication. Unfortunately, medication overuse can make migraines more frequent and severe. It is therefore essential to identify those most at risk of migraine medication overuse, as these are the very people most in need of new, more effective, potentially personalized, migraine medications.
Fiorella Guadagni of the BioBIM (InterInstitutional Multidisciplinary Biobank), San Raffaele Roma Open University, Italy, and colleagues, adopted a machine learning approach based on Multiple Kernel Learning (MKL) to predict the risk of medication overuse in migraine patients. The benefit of MKL is that it provides insight into the machine “reasoning”, rather than simply providing an output, which is seldom the case with other machine learning statistical packages used in the medical sphere. The results were published in the Computational and Structural Biotechnology Journal.
Similar to, but slightly higher than in previous studies, 21% of the 777 enrolled patients self-reported migraine medication overuse lasting at least 2 years. The best performing MKL achieved an AUC of 0.79, which is a measure of accuracy, a score of 1 being most accurate, whereas a standard support vector machine achieved 0.71. The best MKL model was weighted on treatment details and lifestyle-related triggers. When the best MKL models were combined into a single predictor an AUC of 0.83 was achieved in the test dataset.
In the future, the best combination predictor could be used by neurologists to understand which of their patients is most at risk of migraine medication overuse, based on their electronic health record, and tailor advice and or prescriptions accordingly.
That said, the machine learning was limited by the relatively small sample sizes. Further validation in multicenter prospective studies could improve the AUC and is needed before the promising approach can be incorporated into clinical practice.
“This [machine learning technique] is particularly appealing in a context of predictive medicine, in which attributes, routinely collected in electronic health records, may be all used to design new tools for clinical and therapeutic decision making,” concluded the authors.