Table 4.
Main articles using AI for migraine therapy choose and response prediction.
| Reference (First Author, Year) |
Type of AI | Main Results | Number of Patients/Data Analysed |
|---|---|---|---|
| Chartier et al., 2022 [61] |
Random forest machine learning | Data extracted from patients’ drawings had high accuracy in defining poor surgical outcomes in headache patients. | 131 pain drawings |
| Chiang et al., 2023 [52] | Two-level nested logistic regression model | Effectiveness of abortive drugs was accurately evaluated with this model for triptans, ergots, and anti-emetics. | 10,842,795 migraine attack records extracted from an e-diary smartphone application |
| Ciancarelli et al., 2022 [57] | Artificial neural network called ARIANNA (artificial intelligent assistant for neural network analysis) | ARIANNA accurately predicted the post-treatment MIDAS score after biofeedback treatment in 75%. | 20 women with CM |
| Ferroni et al., 2020 [51] |
SVM and Random Optimisation (RO-MO), logistic regression | RO-MO can accurately predict medication overuse in migraine, taking into consideration clinical, biochemical, drug exposure, and lifestyle (four predictors). By using at least 3 RO-MO, accuracy can be higher than 0.87. | 777 migraine patients |
| Fu et al., 2022 [59] | Leave-one-out cross-validation (LOOCV), SVM, and support vector regression (SVR) | 3650 fMRI features accurately distinguished migraine from HS. 70 features accurately predicted response to transcranial vagal nerve stimulation (tVNS). | 70 EM 70 HS |
| Gonzalez-Martinez et al., 2022 [54] |
Classification algorithms (random forests and hyperparameters) and optimization metric (F1 score) | Independently from clinical and demographical features, AI can accurately predict responses to anti-CGRP therapies. | 712 patients with migraine receiving anti-CGRP therapies |
| Hindiyeh et al., 2022 [50] |
NLP and ML algorithms (F1 score) | Data extracted from EHR were compared to reference standards, and the average F1 score for automated extraction was 90.2% for AI for 11 features, suggesting the possibility of using AI for extracting ‘soft’ outcomes. | 1003 patients 2006 encounters |
| Martinelli et al., 2023 [56] |
Random forest, SVM, artificial neural network (ANFIS and MLP), and fuzzy clustering | AI can efficiently predict responses to BoNT-A in CM and HFEM. Only in HFEM a pattern of clinical features can predict responsiveness to BoNT-A. | 113 CM 32 HFEM |
| Matin et al., 2022 [58] | Network and in silico analysis of differential gene expression, using STRING 11.0 database | Aerobic exercise combined with vitamin B12 and magnesium supplementation significantly ameliorated MIDAS and headache features, paralleled by a decline in CGRP levels. | 60 CM |
| Parrales Bravo et al., 2019 [55] | Feature subset selection (C4.5, WrapperSubsetEval, and ClassifierSubsetEval), simulated annealing method (SA), and random tree | AI can predict responsiveness to BoNT-A with an accuracy ranging from 85% (using clinical data) to 91% (HIT-6). | 173 CM |
| Stubberud et al., 2022 [53] | NLP, causal multi-task Gaussian process model, and logistic regression model | AI can help choose the right individual preventive therapy quicker. | 1446 CM |
Abbreviations: AI = artificial intelligence; CGRP = calcitonin gene-related peptide; CM = chronic migraine; EHR = electronic health record; EM = episodic migraine; HFEM = high-frequency episodic migraine; HS = healthy subject; ML = machine learning; NLP = natural language processing; and SVM = support vector machine.