The director of the Neuroinformatics Program at the University of California, Irvine, discussed complexities with different machine learning algorithms in migraine research, including findings from the AMPP study. [WATCH TIME: 5 minutes]
WATCH TIME: 5 minutes
"The reality is that based on the data you have, you have to work with it and find the best models that fit that type of data. Not every model will fit every data set. But also, this causes the problem of reliability of models and validation of those models."
Based on research, there is substantial variability in the clinical manifestations of migraine, both between and within individuals, from attack to attack, and over time. The American Migraine Prevalence and Prevention Study (AMPP), a longitudinal, population-based study of individuals with self-reported headache, conducted annually between 2005 and 2009, collected data on demographics migraine features, disability and depression. At the recently concluded American Headache Society (AHS) Annual Meeting, held June 15-18, in Austin, Texas, investigators presented findings from the study, identifying naturally occurring clusters using a data-driven approach.
Led by Ali Ezzati, MD, the analysis used a self-organizing map (SOM), an unsupervised machine learning technique that preserves topological features of data and is based on competitive learning of a 2-layer artificial neural network. The sample comprised of 4423 individuals, mostly female (83.5%), with an average age of 46.8 years old. Using outcomes of migraine symptom severity score, cutaneous allodynia, monthly headache days, Migraine Disability Assessment Scale, and Patient Health Questionnaire-9, the analysis identified 5 main clusters of migraine, with varying levels of severity.
Following his presentation, Ezzati, director of the Neuroinformatics Program at the University of California, Irvine, sat down with NeurologyLive® to discuss the different big data approaches starting to be incorporated in clinical trials, and whether there are some that offer greater advantages. In addition, he broke down the different clusters observed from the study, and the characteristics that helped determine patient groups.