Table 2.

Main articles using AI for migraine biomarkers identification.

Reference (First Author, Year)	Type of AI	Main Results	Number of Patients/Data Analysed
Akben, 2012 [46]	Multi-layer perceptron neural network	4 Hz of flash stimulation frequency is the most effective frequency, and an 8 s period is necessary to identify migraine at the beta band on the EEG T5-T3 channel	15 migraine patients and 15 HS
Chen, 2022 [47]	Linear discriminant analysis and quadratic discriminant analysis	The change of hemodynamic signals of HS was smaller, while there was a large difference among migraine patients	34 subjects (13 HS, 9 CM, and 12 MOH)
Mitrovic, 2023 [43]	Several models, the best being linear discriminant analysis	The thickness of the left temporal pole, right lingual gyrus, and left pars opercularis was found as markers for MwA classification; the thickness of left pericalcarine gyrus and left pars opercularis was proposed as the features for the classification between MwA-S and MwA-C	78 subjects, among which 46 MwA (22 MwA-S and 24 MwA-C) and 32 HS, with 340 different features used
Tu et al., 2020 [44]	Recursive feature elimination + SVM	Different rsFC can accurately differentiate migraine by HS. No difference in this connectome was detected between MwoA and chronic pain patients. These markers helped to predict response to acupuncture.	144 subjects, among which 70 MwoA, 46 HS, 17 CLBP, and 11 FM

Abbreviations: CLBP = chronic low back pain; CM = chronic migraine; FM = fibromyalgia; HS = healthy subject; MwA = migraine with aura; MwA-S = with simple (i.e., visual) aura; MwA-C = with complex (i.e., different or additional neurological symptoms) aura; MwoA = migraine without aura; rsFC = resting-state functional connectivity; and SVM = support vector machine.