Antonios Danelakis, Håkon Kvisle Abildsnes, Fahim Faisal, Marte‐Helene Bjørk, Dominic Giles, Knut Hagen, Tjaša Kumelj, Manjit Matharu, Parashkev Nachev, Erling Tronvik, Bendik S. Winsvold, Anker Stubberud, Verneri Anttila, Ville Artto, Andrea C. Belin, Anna Bjornsdottir, Gyða Björnsdóttir, D.I. Boomsma, Sigrid Børte, Mona A. Chalmer, Daniel I. Chasman, Bru Cormand, Ester Cuenca-León, George Davey-Smith, Irene de Boer, Martin Dichgans, Tõnu Esko, Tobias Freilinger, Padhraig Gormley, Lyn R. Griffiths, Eija Hämäläinen, Thomas F. Hansen, Aster V. E. Harder, Heidi Hautakangas, Marjo Hiekkala, Maria G. Hrafnsdottir, M. Arfan Ikram, Marjo‐Riitta Järvelin, Risto Kajanne, Mikko Kallela, J Kaprio, Mari Kaunisto, Lisette J. A. Kogelman, Espen S. Kristoffersen, Christian Kubisch, Mitja Kurki, Tobias Kurth, Lenore Launer, T Lehtimäki, Davor Lessel, Lannie Ligthart, Sigurður H. Magnússon, Rainer Malik, Bertram Müller-Myhsok, Carrie Northover, Dale R. Nyholt, Jes Olesen, Aarno Palotie, Priit Palta, Linda M. Pedersen, Nancy Pedersen, Matti Pirinen, Daniëlle Posthuma, Patricia Pozo-Rosich, Alice Pressman, Olli Raitakari, Caroline Ran, Gudrun R. Sigurdardottir, Hreinn Stefánsson, Kari Stefansson, Ólafur Á. Sveinsson, Gisela M. Terwindt, Thorgeir E. Thorgeirsson, Arn M. J. M. van den Maagdenberg, CM van Duijn, Maija Wessman, Bendik S. Winsvold, John-Anker Zwart
BACKGROUND AND OBJECTIVES: In the absence of biomarkers, the true biological footprint of migraine remains incompletely understood. It could perhaps be best characterized using machine learning models of multimodal data. The aim of this study was to (1) develop diagnostic models of migraine using multimodal data and (2) identify data-driven migraine phenotypes. METHODS: This was a cross-sectional machine learning analysis of demographics, self-reported clinical and headache data, and genome-wide genotype data from the Trøndelag Health Study (data collected 1995-1997 and 2006-2008). All participants who were genotyped and completed the headache questionnaire were included. First, predictive machine learning models were developed using genotype data and general clinical data (excluding headache data) to diagnose individuals with migraine vs headache-free controls. Models were optimized on a training set and evaluated on a held-out test set, scored with the area under the receiver operating characteristic curve (AUC). Second, unsupervised models were trained on the headache data and the most predictive features from the diagnostic models to identify subgroups. The subgroups were compared using genome-wide association analyses, conventional polygenic risk scores (PRSs), and machine learning-based genetic risk scores. RESULTS: A total of 43,197 individuals were included in the diagnostic models, and 12,185 individuals were included in the data-driven phenotyping (mean [SD] age 49.1 [16.7] years; 51.7% women). The top-performing diagnostic model was a light gradient boosting machine, with a test set AUC of 0.80 (95% CI 0.78-0.81). Two main clusters were identified, one with 1,425 individuals, 94% of whom met diagnostic criteria for migraine, and another with 10,760 individuals, whereof 71% had nonmigraine headaches. The former was subclustered into 4 relatively distinct groups: one with only men, one with prominent neck pain, one with more musculoskeletal pain, anxiety and depression, and one with "classic" migraine. The groups were better discriminated by machine learning-based genetic risk scores compared with PRSs. DISCUSSION: Migraine can accurately be diagnosed from nonheadache data, suggesting that it is biologically describable by combinations of clinical, genetic, and environmental data. Data-driven phenotyping with such data identifies migraine subgroups with distinct phenotypic and genotypic signals, possibly not captured by current diagnostic criteria-but with potential implications for management.